fdu_core/

lib.rs

//! **fdu** — a fast, incremental file roll-up engine.
//!
//! fdu answers, for any directory in a tree: how big is it, how many files does it hold,
//! what changed most recently, and what kinds of files live in it — hierarchically, for
//! every directory at once, from a single walk.
//!
//! # The shape: three artifacts, one contract
//!
//! 1. **The index** ([`Index`]) — the in-memory hierarchical structure: entry
//!    records plus per-directory roll-up state.
//! 2. **The snapshot** ([`snapshot`]) — that index, serialized.
//! 3. **The change contract** ([`Observation`] and [`Commit`]) —
//!    producers submit verified observations; the index commits clocked effective
//!    changes.
//!
//! Everything else is a producer of observations or a consumer of exact commits. The
//! walker establishes a baseline from upsert observations; the reconciler submits the
//! conditional diff between indexed state and reality; the watch layer submits verified,
//! coalesced observations. The index arbitrates them and re-rolls its reducers; a change
//! feed consumes exact effective changes and state transitions from [`Commit`].
//!
//! A deliberate consequence: **watching is not tied to the roll-up logic.** The index
//! knows `apply(Observation)` and nothing about filesystem events, so a batch scan, a test
//! feeding synthetic observations, and a live watcher are indistinguishable to it.
//!
//! # Freshness is a ladder, not a set of alternatives
//!
//! [`open`] is the conservative, blocking entry point: it loads a compatible snapshot,
//! reconciles the configured filesystem scope, and only then returns. It does not serve
//! the loaded baseline concurrently. Applications that want that model can own an
//! [`IndexHandle`], call the applying reconciliation APIs, and inspect [`Freshness`]
//! while readers continue between short write batches. With the `watch` feature,
//! `watch::Watcher::apply_next` verifies event hints and closes invalidations through
//! subtree reconciliation; neither `open` nor the Python binding starts it implicitly.
//! [`OpenedIndex`] is the additive long-lived owner: its clones share one live identity,
//! cancellation domain, index, and joined shutdown. A cloned [`Index`] remains a
//! detached image and never inherits that authority.
//!
//! ```no_run
//! use fdu_core::{CachePolicy, open};
//! use fdu_core::query::{Basis, Delivery, Scope};
//! use std::path::Path;
//!
//! let basis = Basis { root: Path::new(".").into(), scope: Scope::default(), content: Default::default() };
//! let (index, report) = open(&basis, &Delivery::new(CachePolicy::Auto, None))?;
//! let total = index.total();
//! println!("{} files, {} bytes ({:?})", total.files, total.bytes, report.path_taken);
//! # Ok::<(), fdu_core::Error>(())
//! ```
//!
//! # Build features
//!
//! - `watch` — the OS-native watch layer.
//!
//! `fdu-core` has no default build features. The command and Python packages opt into
//! watch, while embedding consumers can retain the smaller one-shot engine.
//!
//! `.gitignore` handling is not a build feature: it has no dependency, so it is always
//! compiled in, and whether a scan reads control files is decided at runtime by
//! [`ScanConfig::read_controls`].

pub mod admission;
pub mod cache;
pub mod classify;
pub mod content;
pub mod control;
pub mod counters;
mod emit;
mod engine_contract;
mod execution;
mod index;
mod opened;
mod platform_tuning;
mod progress;
pub mod query;
pub mod scan;
pub mod snapshot;
mod stored_state;
#[cfg(test)]
mod test_support;

/// The crate README's Rust examples, compiled and run as doctests.
///
/// crates.io shows the README as this crate's front page, so its examples are the first
/// code a reader copies. Nothing compiled them, and they kept naming an `AnalysisProfile`
/// type for a release after the content axis became `AnalysisSet`.
#[cfg(doctest)]
#[doc = include_str!("../README.md")]
pub struct ReadmeDoctests;

// Ungated: rendering is not a command-line concern. It was behind `cli` only because it
// took its ANSI colour types from clap, so the library could produce a report and not
// print it -- and a display note added elsewhere on this branch called into here and
// broke the no-default-features build, which is that gap showing itself.
pub mod report_format;

#[cfg(feature = "watch")]
pub mod watch_session;

#[cfg(feature = "watch")]
pub mod watch;

#[cfg(feature = "watch")]
pub use crate::watch_session as session;

pub use crate::admission::HiddenPolicy;
pub use crate::cache::{
    CachePaths, CacheScope, CacheState, CacheStatus, ClearSummary, ContentInfo, ContentState,
    ContentStatus, LeftoverKind, SnapshotInfo, StaleReason, cache_status, clear_all_caches,
    clear_cache, list_caches,
};
pub use crate::control::{
    CONTROL_FILE_NAME, ControlAdmission, ControlCoverage, ControlIdentity, ControlLimits,
    ControlMatcher, ControlObservation, ControlRefusalReason, ControlTable, DEFAULT_CONTROL_BUDGET,
    DEFAULT_CONTROL_LINE_LIMIT, RefusedControl, is_control_file,
};
pub use crate::engine_contract::{
    Attrs, ChangeOutcome, ChangePoll, ChangeRequest, Clock, Commit, ContinuationId, CountResult,
    Coverage, CoverageReason, DEFAULT_COUNT_CAP, DiscoveryProgress, EffectiveChange, EngineVersion,
    EntryKind, EntryValue, Error, Expectation, Fingerprint, FlatPage, Freshness, Impact,
    ImpactDomain, IndexState, InvalidateReason, Issue, IssueKind, IssueSummary, Knowledge,
    LifecyclePhase, LimitedProjection, MAX_CONTINUATION_RECORD_BYTES, MAX_COUNT_CAP,
    MAX_DIRTY_PATHS, MAX_ISSUE_MESSAGE_BYTES, MAX_ISSUE_PATH_BYTES, MAX_PAGE_ROWS, MAX_PAGE_WORK,
    MAX_READ_PROJECTIONS, MAX_REPORT_VIEWS, MAX_RETAINED_ISSUES, MIN_JOURNAL_CAPACITY_BYTES,
    Observation, ObservationOp, Op, PageRequest, PathExpectation, PathState, PortablePath,
    ProjectionRefusal, ProjectionResult, Provenance, QueryLimit, ReadDiagnostics, ReadProjection,
    ReadRequest, ReadResponse, RefreshRejection, RefreshResult, RejectedRefreshPath, ReportRequest,
    Result, RowShape, ScanScope, SemanticIdentity, SessionId, Source, StateTransition, Status,
    TreePage, Work,
};
pub use crate::index::{
    ApplyOutcome, ApplyStats, ChildSnapshot, DEFAULT_JOURNAL_CAPACITY_BYTES, EntryId, ExtTally,
    Index, IndexHandle, PartitionRollUp, PartitionRollUpSummary, RollUp, RollUpSummary, Since,
};
pub use crate::opened::{
    DiscoveryBudget, MAX_PRIORITY_PATHS, MAX_REFRESH_PATHS, OpenOptions, OpenedIndex,
};
// Ungated with report_format, for the same reason: one-shot planning is an execution
// strategy, not a front end. A caller wanting one report without retaining an index was
// previously required to compile the command line to get it (fdu-z7sp).
pub use crate::execution::{
    Load, OutcomeClass, PerformanceSummary, Plan, Route, Verify, plan, prepare_report,
    prepare_report_with_progress, prepare_report_with_scan_diagnostics, throughput_rates,
};
pub use crate::progress::{Progress, ProgressPhase, ProgressSnapshot};
pub use crate::scan::{ReconcileReport, ScanConfig, ScanOrder, ScanReport};
pub use crate::stored_state::{
    AnalyzerProvenance, ContentAdmission, ContentTierIdentity, ControlTierIdentity, EntryScope,
    EntryTierIdentity, Serves, SnapshotIdentity, serves_snapshot,
};
#[cfg(feature = "watch")]
pub use crate::watch_session::{Batch, Change, ChangeKind, SaveOutcome, Session};

use crate::execution::{Admission, RunFacts, SaveTargets, StoreHeader};
use std::ffi::OsString;
use std::path::{Path, PathBuf};

/// How to open a tree.
#[cfg(test)]
#[derive(Clone, Debug, Default)]
pub(crate) struct OpenFixture {
    /// Walk settings.
    pub scan: ScanConfig,
    /// Where the snapshot for this root lives.
    ///
    /// `None` disables the cache regardless of policy, which is what a caller with no
    /// writable cache directory gets.
    pub cache_path: Option<PathBuf>,
    /// How the snapshot may be used.
    pub policy: CachePolicy,
    /// Answer from the snapshot alone.
    pub stale_ok: bool,
    /// Optional streaming content analysis. Disabled preserves metadata-only behavior.
    pub analysis: content::AnalysisRequest,
}

#[cfg(test)]
impl OpenFixture {
    /// Compose model inputs for a unit-test fixture.
    pub fn split(&self, root: impl Into<PathBuf>) -> (query::Basis, query::Delivery) {
        (
            query::Basis {
                root: root.into(),
                scope: self.scan.clone().into(),
                content: self.analysis.profile,
            },
            query::Delivery {
                cache: self.policy,
                stale_ok: self.stale_ok,
                cache_path: self.cache_path.clone(),
                accept_partial: false,
                watch: None,
                workers: query::Workers {
                    scan: self.scan.threads,
                    analysis: self.analysis.workers,
                },
                batch_size: self.scan.batch_size,
                order: self.scan.order,
            },
        )
    }
}

#[cfg(test)]
pub(crate) fn open_fixture(root: &Path, config: &OpenFixture) -> Result<(Index, OpenReport)> {
    let (basis, delivery) = config.split(root);
    open(&basis, &delivery)
}
#[cfg(test)]
pub(crate) fn open_fixture_with_pending_save(
    root: &Path,
    config: &OpenFixture,
) -> Result<(std::sync::Arc<Index>, OpenReport, PendingSave)> {
    let (basis, delivery) = config.split(root);
    open_with_pending_save(&basis, &delivery)
}

/// Whether a request may read and write the snapshot cache.
///
/// Three values, because the question a caller has is whether fdu should decide, keep
/// the cache, or stay out of it. What `Auto` does depends on the analysis: the planner
/// reads and writes only where a later request can use what it stores, so the default
/// never pays for a store that nothing reads. Which paths read and write under each
/// value is [`crate::execution::plan`]'s decision and the cache design's policy table.
///
/// Answering from the snapshot without touching the tree is a separate choice,
/// [`query::Delivery::stale_ok`], since it changes what the answer promises rather than
/// what the run stores.
#[derive(Clone, Copy, PartialEq, Eq, Debug, Default)]
pub enum CachePolicy {
    /// Read and write where it pays for this analysis.
    ///
    /// A one-shot metadata report neither reads nor writes: revalidating a snapshot stats
    /// every entry, as a cold walk does, and no later one-shot report reads what it would
    /// write. Content analysis reads and writes the snapshot and its content sidecar,
    /// because the sidecar spares re-reading unchanged files. An [`open`] session, a
    /// watch, and a refresh read, revalidate, and write, because the session is itself the
    /// later reader.
    ///
    /// A root has one cache path, and its snapshot carries the scan scope that wrote it.
    /// A read under another scope normally treats that snapshot as absent and scans cold.
    /// The lawful exception is a controls-off request with the same entry identity: it
    /// projects a controls-on snapshot into the requested blind scope and never replaces
    /// the stronger image with that projection.
    ///
    /// Every default request observes `.gitignore` control state -- the one-shot
    /// `fdu <dir>` and [`prepare_report`], `fdu --watch <dir>`, and a default [`open`] --
    /// so they share one scope: a watch or a report that reads the snapshot, as content
    /// analysis does, starts warm from a session's snapshot or an `On` report's. A request
    /// that turns [`ScanConfig::read_controls`] off is a second scope, but every route can
    /// start from a default snapshot by discarding its control tier while loading.
    #[default]
    Auto,
    /// Read and write where `Auto` does, and also write after a one-shot report.
    ///
    /// The way to leave a current snapshot behind a one-shot report, for a later
    /// [`query::Delivery::stale_ok`] answer or a warm session. A summary that would
    /// otherwise retain nothing builds the index it writes. Reads are the same as `Auto`'s:
    /// loading a snapshot that a full revalidation then re-stats costs more than a cold
    /// walk, and caching never changes an answer, so there is nothing to gain by forcing it.
    On,
    /// Ignore any snapshot and leave nothing behind.
    Off,
}

impl CachePolicy {
    /// Whether this policy may read an existing snapshot on some route.
    fn reads(self) -> bool {
        matches!(self, Self::Auto | Self::On)
    }
}

/// Which tier of the freshness ladder an [`open`] actually used.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub enum OpenPath {
    /// No usable snapshot: the tree was walked from scratch.
    ColdScan,
    /// A snapshot was loaded and reconciled against the filesystem.
    WarmRevalidate,
    /// A snapshot answered on its own; the filesystem was never consulted.
    ///
    /// The only tier that can be stale, and it says so rather than implying currency.
    CacheOnly,
}

/// Live entries past which a one-shot index is released off the caller's thread.
///
/// Releasing an index frees every entry's name and every directory's child list, one
/// allocation at a time, and nothing reads the result: on a million-entry Linux tree it
/// was 95 ms of a 1.39 s `--cache off` report, all of it after the answer was complete
/// (exp-160). At that rate this threshold is about 6 ms of release, well above the tens
/// of microseconds a thread spawn costs; smaller indexes release inline, so a small
/// report never starts a thread to save almost nothing.
const BACKGROUND_RELEASE_MIN_ENTRIES: u64 = 64 * 1024;

/// Drop one reference to a one-shot index, moving the final release off this thread.
///
/// Only the last reference does any work; every other one is an ordinary decrement.
/// `Arc::into_inner` makes that decision race-free when the caller and a snapshot
/// writer let go concurrently: exactly one of them receives the index. A large index
/// then goes to a detached, named thread. That thread holds no engine state and reports
/// nothing — its only effect is returning memory — so there is nothing to join: a
/// process that exits first lets the operating system reclaim the pages instead, and a
/// long-lived caller gets the memory back moments later rather than before its answer.
/// A host that cannot spawn the thread releases the index inline, as before.
///
/// With counters on, the release stays inline. A thread's counts reach the totals only
/// when it exits, so frees on a detached thread would land in a run's report or miss it
/// depending on timing; counting runs trade the saving for a deterministic record.
///
/// On Windows the release stays inline too. `ExitProcess` terminates other threads
/// without notice, so a release still running at exit can die holding the process heap's
/// lock while DLL detach code allocates, and the saving was never measured there.
pub(crate) fn release_index(index: std::sync::Arc<Index>) {
    let Some(index) = std::sync::Arc::into_inner(index) else {
        return;
    };
    if index.len() < BACKGROUND_RELEASE_MIN_ENTRIES || crate::counters::enabled() || cfg!(windows) {
        return;
    }
    let spawned = std::thread::Builder::new()
        .name("fdu-index-release".to_string())
        .spawn(move || drop(index));
    // On failure the builder drops the closure, and with it the index, on this thread.
    drop(spawned);
}

/// A snapshot write running alongside rendering.
///
/// The index is read-only by the time this starts, so the writer and the renderer are
/// two readers of the same data. The handle exists so the process can join before it
/// exits: an abandoned write would leave a half-written snapshot for the next run to
/// reject, turning a warm start into a cold one for no reason.
#[derive(Debug)]
#[must_use = "join the save before exiting or the snapshot may be abandoned"]
pub struct PendingSave {
    workers: Vec<(&'static str, std::thread::JoinHandle<Result<()>>)>,
}

impl PendingSave {
    /// Nothing to wait for.
    pub(crate) fn none() -> Self {
        Self { workers: Vec::new() }
    }

    /// Whether the metadata snapshot is among the writes being waited for.
    pub(crate) fn writes_metadata(&self) -> bool {
        self.workers.iter().any(|(name, _)| *name == "metadata")
    }

    /// Wait for the write to finish, returning its result.
    ///
    /// A failed save is the caller's to report, not to die on: the answer already
    /// rendered is still correct, and only the next run's warmth is lost.
    pub fn join(mut self) -> Result<()> {
        let mut first_error = None;
        for (name, worker) in self.workers.drain(..) {
            let outcome = worker
                .join()
                .unwrap_or_else(|_| Err(Error::Snapshot(format!("{name} cache writer panicked"))));
            if first_error.is_none() {
                first_error = outcome.err();
            }
        }
        first_error.map_or(Ok(()), Err)
    }
}

impl Drop for PendingSave {
    fn drop(&mut self) {
        // A dropped handle still waits: losing the write silently would be worse than
        // the brief delay, and this only happens on a path that forgot to join.
        for (_, worker) in self.workers.drain(..) {
            let _ = worker.join();
        }
    }
}

/// What [`open`] did.
#[derive(Debug)]
pub struct OpenReport {
    /// Cache tier used to produce the returned index.
    pub path_taken: OpenPath,
    /// Filesystem walk results, including any partial errors.
    pub scan: ScanReport,
    /// Content-analysis work performed after metadata reconciliation.
    pub analysis: Option<content::AnalysisReport>,
    /// Reusable records restored from the independently versioned content sidecar.
    pub content_cache: content::ContentCacheLoad,
    /// Whether the returned index was projected from a stronger controls-on snapshot.
    pub projected: bool,
}

impl OpenReport {
    /// Whether every path in the requested scan scope was read successfully.
    pub fn is_complete(&self) -> bool {
        self.scan.is_complete()
            && self.analysis.as_ref().is_none_or(content::AnalysisReport::is_complete)
    }

    /// Per-path errors that make this result partial.
    pub fn errors(&self) -> &[Error] {
        &self.scan.errors
    }

    /// Human-readable diagnostics for every operational condition that makes this result partial.
    pub fn error_messages(&self) -> Vec<String> {
        let mut errors = self.scan.errors.iter().map(ToString::to_string).collect::<Vec<_>>();
        if let Some(message) =
            self.analysis.as_ref().and_then(content::AnalysisReport::failure_message)
        {
            errors.push(message);
        }
        errors
    }
}

/// Open a tree, using the snapshot cache when one is usable.
///
/// On the warm path the snapshot is loaded and then reconciled against the filesystem
/// before being returned. Errors are represented as partial freshness and the previous
/// complete snapshot is left untouched; callers must inspect [`OpenReport::is_complete`]
/// or [`Index::freshness`] before treating totals as complete.
///
/// The index observes control state as [`ScanConfig::read_controls`] says, and the
/// default is on: the index exposes [`Index::controls`] and [`Index::is_ignored`], and a
/// watch over it maintains them. A one-shot report from [`prepare_report`] observes it on
/// the same terms, so a default `open` and a default report share one snapshot scope and
/// each starts warm from the other's snapshot. A caller wanting a single answer should use
/// [`prepare_report`].
///
/// A caller that reads no ignore classification may turn the field off. Its `open` reads
/// no `.gitignore`, and its index answers [`Index::is_ignored`] and [`Index::controls`]
/// with [`Error::ControlStateNotObserved`] rather than calling every entry unignored. Its
/// snapshot is of another scope. When its entry identity matches a default snapshot, the
/// loader discards that snapshot's control tier and constructs the returned index directly
/// in the requested controls-off scope. The projected index never replaces the stronger
/// snapshot, including during a watch session.
pub fn open(basis: &query::Basis, delivery: &query::Delivery) -> Result<(Index, OpenReport)> {
    let (index, report, pending) = open_with_pending_save(basis, delivery)?;
    // Joining first is what makes the unwrap infallible: the writer held the only other
    // reference, and this is the blocking entry point, so by here it has finished and
    // dropped it. `try_unwrap` rather than a clone keeps the owned-`Index` signature
    // honest — a fallback clone here would quietly reintroduce the copy the shared
    // writer exists to avoid.
    let wrote_metadata = pending.writes_metadata();
    pending.join()?;
    let mut index = std::sync::Arc::into_inner(index)
        .expect("the joined writer released the only other reference");
    // Only here, after the join succeeded: a write that failed leaves the debt on the
    // index, and a caller of [`open_with_pending_save`] who keeps the index through a
    // failed write keeps the debt with it, so its next writing pass writes.
    if wrote_metadata {
        index.set_persistence_owed(false);
    }
    Ok((index, report))
}

/// Open a tree, returning the snapshot write for the caller to join.
///
/// The blocking [`open`] is the right default; a caller that renders its own output can
/// use this to overlap the write with rendering and join before exiting.
///
/// This path always loads a usable snapshot, because its callers — live sessions and
/// library consumers holding the index — amortise the load across everything they do
/// with it. A one-shot report cannot; the internal report planner decides per request
/// whether the read pays and routes through the gated variant below.
pub fn open_with_pending_save(
    basis: &query::Basis,
    delivery: &query::Delivery,
) -> Result<(std::sync::Arc<Index>, OpenReport, PendingSave)> {
    let mut query = query::Query::default();
    query.selection.ignored = basis.scope.population;
    let request = query::Request::new(basis.clone(), query, std::time::SystemTime::now());
    let plan = plan(&request, delivery, Route::Retained).map_err(Error::InvalidRequest)?;
    execute(&plan, &request.basis, false, None)
        .map(|(index, report, pending, _diagnostics)| (index, report, pending))
}

/// Reverify a retained index, refresh requested content, and persist according to delivery.
///
/// A partial pass retains its verified facts and may save only the content tier beside
/// a compatible complete snapshot. The returned report describes metadata changes.
///
/// The metadata write is owed by the index, not by the pass: a pass that mutated the
/// entry tier without writing it -- because it was partial, or its policy does not write
/// -- leaves the index holding facts the snapshot lacks, and the next complete pass under
/// a writing policy writes them even when it changed nothing itself. Only a completed
/// write clears that debt, so a failed one is retried by the next pass as well.
pub fn refresh(
    index: &mut Index,
    basis: &query::Basis,
    delivery: &query::Delivery,
) -> Result<ReconcileReport> {
    let mut query = query::Query::default();
    query.selection.ignored = basis.scope.population;
    let request = query::Request::new(basis.clone(), query, std::time::SystemTime::now());
    let plan = plan(&request, delivery, Route::Refresh).map_err(Error::InvalidRequest)?;
    validate_basis_root(index.root_path(), basis)?;
    request.validate_read(&query::Basis::held_by(index)).map_err(Error::InvalidRequest)?;
    let scan = basis.scope.scan_config(delivery);
    scan.validate_for_scope(index.scope())?;
    // `plan` refuses the one policy that verifies nothing on this route, so there is no
    // second refusal here to keep in step with it.
    debug_assert_eq!(plan.verify(), Verify::Filesystem, "a refresh plan verifies the tree");
    let report = scan::reconcile(index, &scan, &mut |_| {})?;
    let cached = load_content(index, basis, delivery)?;
    let analysis = basis.content.is_enabled().then(|| {
        content::analyze_index(
            index,
            content::AnalysisRequest { profile: basis.content, workers: delivery.workers.analysis },
        )
    });
    if report.apply.mutated() {
        index.set_persistence_owed(true);
    }
    let written = persist_index_changes(
        index,
        &plan,
        index.persistence_owed(),
        !cached.usable
            || cached.stale > 0
            || analysis.as_ref().is_some_and(|report| report.applied > 0),
    )?;
    if written {
        index.set_persistence_owed(false);
    }
    Ok(report)
}

pub(crate) fn validate_basis_root(held: &Path, basis: &query::Basis) -> Result<()> {
    let requested = basis.root.canonicalize().map_err(|error| Error::io(&basis.root, error))?;
    let held = held.canonicalize().map_err(|error| Error::io(held, error))?;
    if requested != held {
        return Err(Error::InvalidRequest(query::RequestError::RootMismatch { held, requested }));
    }
    Ok(())
}

/// Why a stale answer has no snapshot to answer from, and what recovers.
///
/// [`query::Delivery::stale_ok`] is the one delivery that cannot fall back to a scan, so its failure
/// is the only place a caller learns that the snapshot is missing or of another scope.
/// Type rules determine each file's category, so attaching another registry would make a
/// cache-only answer false. Control state is another common difference: a snapshot
/// without it was written by a request that turned observation off, or by a
/// release from before observation was the default, and a default cache-only request after
/// it would otherwise fail with no hint that a snapshot exists at all. Control limits: both
/// scopes observe, and their identities are hashes, so "a different scan scope" would
/// describe a request whose only difference is a limit the caller chose and can choose
/// again.
fn unusable_snapshot_message(refused: Option<SnapshotIdentity>, wanted: &ScanConfig) -> String {
    // `on`, not `auto`: a one-shot metadata report under `auto` writes nothing, so that
    // remedy would fail again for exactly the query that asked. `on` writes after every
    // complete run, building an index for a summary that would otherwise retain none.
    const PREFIX: &str = "no usable snapshot for this root and scan scope";
    const NEVER_SCANS: &str = "a stale answer never scans";
    const VERIFIED: &str = "ask for a verified answer, which scans when none serves";
    const LEAVE_ONE: &str = "run the request once with the `on` cache policy to leave one";
    let wanted_scope = wanted.scope();
    let differs_only_in_ignore_rules = |stored: ScanScope| {
        ScanScope { ignore_rules_fingerprint: wanted_scope.ignore_rules_fingerprint, ..stored }
            == wanted_scope
    };
    let Some(refused) = refused else {
        return format!("{PREFIX}; {NEVER_SCANS}, so {LEAVE_ONE}, or {VERIFIED}");
    };
    let stored = refused.scan_scope();
    if stored.type_rules_fingerprint != wanted_scope.type_rules_fingerprint {
        return format!(
            "{PREFIX}: the cached snapshot was taken under different file type rules; \
             {NEVER_SCANS}, so use the snapshot's type registry, {LEAVE_ONE} under this request's rules, or {VERIFIED}"
        );
    }
    if differs_only_in_ignore_rules(stored) {
        // Named without a knob, because the command line and the library spell the switch
        // differently and this message is the engine's.
        if stored.ignore_rules_fingerprint == 0 {
            return format!(
                "{PREFIX}: the cached snapshot has no .gitignore state, because the request \
                 that wrote it did not observe it, and this request does; {NEVER_SCANS}, so \
                 {LEAVE_ONE} for this scope, {VERIFIED}, or turn .gitignore observation off as \
                 that request did"
            );
        }
        if wanted_scope.observes_controls() {
            let stored_limits = match refused.controls {
                ControlTierIdentity::Observed { limits } => limits,
                ControlTierIdentity::NotObserved => crate::control::ControlLimits::default(),
            };
            let changed = changed_control_limits(stored_limits, wanted.control_limits);
            if !changed.is_empty() {
                return format!(
                    "{PREFIX}: the cached snapshot was taken under other .gitignore limits \
                     ({changed}); {NEVER_SCANS}, so repeat the request with the snapshot's \
                     limits, {LEAVE_ONE} for these limits, or {VERIFIED}"
                );
            }
        }
    }
    format!(
        "{PREFIX}: the cached snapshot has a different scan scope; {NEVER_SCANS}, so \
         {LEAVE_ONE} for this scope, or {VERIFIED}"
    )
}

/// Each control limit that differs between a snapshot and a request, both values named.
fn changed_control_limits(
    stored: crate::control::ControlLimits,
    wanted: crate::control::ControlLimits,
) -> String {
    let display = crate::control::limit_display;
    let changed: Vec<String> = [
        ("budget", stored.budget, wanted.budget),
        ("line limit", stored.line_limit, wanted.line_limit),
    ]
    .into_iter()
    .filter(|(_, stored, wanted)| stored != wanted)
    .map(|(name, stored, wanted)| {
        format!("{name} {}, where this request asks for {}", display(stored), display(wanted))
    })
    .collect();
    changed.join(", and ")
}

/// [`open_with_pending_save`] with the snapshot read under the caller's control.
///
/// `read_snapshot: false` skips loading an existing snapshot and takes the cold-scan
/// path: for a one-shot metadata query, revalidation stats every entry regardless, so
/// the load and the reconciliation against it are additive cost with nothing to
/// amortise them — measured on macOS/APFS over 494,031 entries, warm revalidation cost
/// 4.8 s against 3.6 s for the cold path, whose write-behind the read could at best
/// have saved ~50 ms of. Persistence is decided separately, by [`Plan::persists`] and
/// then per tier by [`Plan::writes`].
///
/// A stale answer reads regardless — for [`query::Delivery::stale_ok`] the snapshot is
/// the contract, not a cost choice.
pub(crate) fn execute(
    plan: &Plan,
    basis: &query::Basis,
    collect_scan_diagnostics: bool,
    progress: Option<&Progress>,
) -> Result<(std::sync::Arc<Index>, OpenReport, PendingSave, Option<scan::ScanDiagnostics>)> {
    let scan_config =
        ScanConfig { progress: progress.cloned(), ..basis.scope.scan_config(plan.delivery()) };
    let analysis_request = content::AnalysisRequest {
        profile: basis.content,
        workers: plan.delivery.workers.analysis,
    };
    let delivery = plan.delivery();
    let root = &basis.root;
    let root = root.canonicalize().map_err(|e| Error::io(root, e))?;
    // Before the snapshot, not at the scan that may never happen: a scope this build cannot
    // honour has no answer at any delivery, and checking it where the scan runs made
    // `--stale-ok` report a snapshot miss for a request every other delivery refuses --
    // which failure a run named then depended on how it was delivered (`refusal-order`).
    scan_config.validate()?;
    let canonical_basis = query::Basis { root: root.clone(), ..basis.clone() };
    // What the store holds, for the plan to admit: the index a served snapshot supplied,
    // and the root and identity the snapshot declares whether or not it served, kept so a
    // policy that cannot scan says why it has no answer rather than only that it has none.
    let mut loaded = None;
    let mut stored: Option<(PathBuf, SnapshotIdentity)> = None;
    if let (Load::Snapshot, Some(cache_path)) = (plan.load(), &delivery.cache_path) {
        if let Some(progress) = progress {
            progress.enter(ProgressPhase::Loading);
        }
        match snapshot::load_serving(
            cache_path,
            scan_config.types_shared(),
            scan_config.snapshot_identity(),
        )? {
            snapshot::LoadOutcome::Served { index, stored: identity } => {
                stored = Some((index.root_path().to_path_buf(), identity));
                loaded = Some(index);
            }
            snapshot::LoadOutcome::Refused { identity, root: stored_root } => {
                stored = Some((stored_root, identity));
            }
            snapshot::LoadOutcome::Absent => {}
        }
    }
    let cache_only = plan.verify() == Verify::None;
    // A policy that cannot scan is admitted on its content tier too, so its sidecar is
    // loaded before the one admission below. A verifying route loads it after reconciling.
    // Deliberately no reconciliation there: that tier never touches the tree, and the
    // index is marked unverified so the answer cannot claim a currency it has not earned
    // -- a snapshot records the freshness it was written with, which was true then.
    let content_cache = match (&mut loaded, cache_only) {
        (Some(index), true) => {
            index.mark_unverified();
            Some(load_content(index, basis, delivery)?)
        }
        _ => None,
    };
    // The one admission of stored state on every route that reads it. A sidecar serves
    // only its own identity, so restoring one record per visited regular file means it
    // holds the complete answer to this request: restore already walked that set, so
    // compare `hits` to files visited, not a second walk and not unique `PathBuf` keys.
    let admission = {
        let header = stored.as_ref().map(|(stored_root, identity)| StoreHeader {
            root: stored_root,
            snapshot: *identity,
            content: loaded
                .as_ref()
                .and_then(Index::content)
                .and_then(|content| content.identity()),
            content_complete: content_cache
                .as_ref()
                .is_some_and(|cache| cache.usable && cache.hits == cache.candidates),
        });
        plan.admit(header.as_ref(), &canonical_basis)
    };

    if cache_only {
        let relation = match admission {
            Admission::Serve(relation) => relation,
            Admission::NoLocation => {
                return Err(Error::Snapshot(
                    "no cache location is configured; configure a cache location before asking for a stale answer".into(),
                ));
            }
            Admission::WrongRoot => {
                return Err(Error::Snapshot(
                    "the configured snapshot belongs to a different root; choose this root's cache location, or run the request once with the `on` cache policy to replace it with this root's snapshot".into(),
                ));
            }
            Admission::Missing => {
                return Err(Error::Snapshot(unusable_snapshot_message(None, &scan_config)));
            }
            Admission::WrongScope => {
                return Err(Error::Snapshot(unusable_snapshot_message(
                    stored.map(|(_, identity)| identity),
                    &scan_config,
                )));
            }
            Admission::IncompleteContent => {
                return Err(Error::Snapshot(
                    "no complete usable content sidecar for this root and analysis profile".into(),
                ));
            }
        };
        let index = loaded.expect("a served admission is of the loaded snapshot");
        let content_cache =
            content_cache.expect("the sidecar was loaded beside the served snapshot");
        return Ok((
            std::sync::Arc::new(index),
            OpenReport {
                path_taken: OpenPath::CacheOnly,
                scan: ScanReport::default(),
                analysis: None,
                content_cache,
                projected: relation == Serves::ProjectControlsOff,
            },
            PendingSave::none(),
            None,
        ));
    }

    if let Admission::Serve(relation) = admission {
        let mut index = loaded.expect("a served admission is of the loaded snapshot");
        let projected = relation == Serves::ProjectControlsOff;
        let reconciled = scan::reconcile(&mut index, &scan_config, &mut |_| {})?;
        let scan_report = reconciled.scan;
        index.establish_baseline();
        let content_cache = load_content(&mut index, basis, delivery)?;
        let analysis = basis
            .content
            .is_enabled()
            .then(|| content::analyze_index_observed(&mut index, analysis_request, progress));
        // A reconciliation that mutated nothing leaves an index that serializes to the
        // bytes already on disk, so rewriting it is pure cost: the clone, the encode,
        // and the write all produce a file identical to the one just read. Each artifact
        // is judged separately because content and metadata are invalidated separately.
        if reconciled.apply.mutated() {
            index.set_persistence_owed(true);
        }
        let facts = run_facts(
            &index,
            basis,
            plan,
            index.persistence_owed(),
            analysis.as_ref().is_some_and(|report| report.applied > 0) || content_cache.stale > 0,
            projected,
            || stored_entries(delivery, &root),
        );
        // The index is shared read-only from here, so the debt a completed write clears
        // is cleared by the blocking [`open`] once it has joined the write.
        let index = std::sync::Arc::new(index);
        let pending = spawn_save(&index, &plan.delivery, plan.writes(facts), progress);
        return Ok((
            index,
            OpenReport {
                path_taken: OpenPath::WarmRevalidate,
                scan: scan_report,
                analysis,
                content_cache,
                projected,
            },
            pending,
            None,
        ));
    }

    let (mut index, scan_report, scan_diagnostics) = if collect_scan_diagnostics {
        let (index, report, diagnostics) =
            scan::scan_into_index_with_diagnostics(&root, &scan_config)?;
        (index, report, Some(diagnostics))
    } else {
        let (index, report) = scan::scan_into_index(&root, &scan_config)?;
        (index, report, None)
    };
    let content_cache = load_content(&mut index, basis, delivery)?;
    let analysis = basis
        .content
        .is_enabled()
        .then(|| content::analyze_index_observed(&mut index, analysis_request, progress));
    let facts =
        run_facts(&index, basis, plan, true, true, false, || stored_entries(delivery, &root));
    let index = std::sync::Arc::new(index);
    let pending = spawn_save(&index, &plan.delivery, plan.writes(facts), progress);
    Ok((
        index,
        OpenReport {
            path_taken: OpenPath::ColdScan,
            scan: scan_report,
            analysis,
            content_cache,
            projected: false,
        },
        pending,
        scan_diagnostics,
    ))
}

/// The entry tier the snapshot at the delivery's cache location declares for `root`.
///
/// Read for the content tier's pairing rule alone, by a route that attempted no load;
/// a route that did reads the header it already holds.
fn stored_entries(delivery: &query::Delivery, root: &Path) -> Option<EntryTierIdentity> {
    delivery
        .cache_path
        .as_ref()
        .and_then(|path| snapshot::read_header(path).ok().flatten())
        .filter(|stored| stored.root == root)
        .map(|stored| stored.identity.entries)
}

fn run_facts(
    index: &Index,
    basis: &query::Basis,
    plan: &Plan,
    entries_changed: bool,
    content_changed: bool,
    projected: bool,
    stored_entries: impl FnOnce() -> Option<EntryTierIdentity>,
) -> RunFacts {
    let entries_verified = stored_state::entries_writable(index);
    let paired_entries = plan.persists()
        && !entries_verified
        && stored_state::content_tier_writable(index, stored_entries);
    RunFacts {
        entries_verified,
        entries_changed,
        content_changed,
        content_requested: basis.content.is_enabled(),
        projected,
        paired_entries,
    }
}

/// Execute the persistence policy for a live index without retaining a lock while writing.
#[cfg(feature = "watch")]
pub(crate) fn persist_index(index: &Index, plan: &Plan) -> Result<bool> {
    persist_index_changes(index, plan, true, true)
}

fn persist_index_changes(
    index: &Index,
    plan: &Plan,
    entries_changed: bool,
    content_changed: bool,
) -> Result<bool> {
    let basis = query::Basis::held_by(index);
    let delivery = plan.delivery();
    if !plan.persists() || delivery.cache_path.is_none() {
        return Ok(false);
    }
    let stored = delivery.cache_path.as_deref().map(snapshot::read_header).transpose()?.flatten();
    // The same admission a load makes, over the header alone: the index's root is the
    // canonical one the header records, and its scope is the plan's, verified against the
    // index before any pass mutated it. Whatever the store cannot serve is replaced.
    let admitted = query::Basis { root: index.root_path().to_path_buf(), ..plan.basis().clone() };
    let header = stored.as_ref().map(|info| StoreHeader {
        root: &info.root,
        snapshot: info.identity,
        content: None,
        content_complete: false,
    });
    let relation = match plan.admit(header.as_ref(), &admitted) {
        Admission::Serve(relation) => relation,
        // The content arm needs a plan that verifies nothing, and no such plan writes.
        Admission::NoLocation
        | Admission::Missing
        | Admission::WrongRoot
        | Admission::WrongScope
        | Admission::IncompleteContent => Serves::Refuse,
    };
    let projected = relation == Serves::ProjectControlsOff;
    let writes = plan.writes(run_facts(
        index,
        &basis,
        plan,
        entries_changed || relation == Serves::Refuse,
        content_changed,
        projected,
        || {
            stored
                .as_ref()
                .filter(|info| info.root == index.root_path())
                .map(|info| info.identity.entries)
        },
    ));
    let Some(path) = delivery.cache_path.as_ref() else {
        return Ok(false);
    };
    if writes.metadata {
        snapshot::save(index, path)?;
    }
    if writes.content {
        content::save_content_cache(index, &content::content_cache_path(path))?;
    }
    Ok(writes.metadata)
}

fn load_content(
    index: &mut Index,
    basis: &query::Basis,
    delivery: &query::Delivery,
) -> Result<content::ContentCacheLoad> {
    let (true, Some(snapshot_path)) = (delivery.cache.reads(), delivery.cache_path.as_deref())
    else {
        return Ok(content::ContentCacheLoad::default());
    };
    let wanted = index.content_identity(basis.content);
    content::load_content_cache(index, &wanted, &content::content_cache_path(snapshot_path))
}

/// Start the cache writes a completed open still needs, each tier under its own rule.
///
/// The snapshot is written only after a complete pass
/// ([`stored_state::entries_writable`]): a snapshot recording a partial view would be
/// served as fact on the next run, and the existing complete snapshot is better than that.
/// The content sidecar keeps the records the pass verified
/// ([`stored_state::content_tier_writable`]); a partial pass may replace it only beside a
/// stored snapshot of the same entry tier.
fn spawn_save(
    index: &std::sync::Arc<Index>,
    delivery: &query::Delivery,
    writes: SaveTargets,
    progress: Option<&Progress>,
) -> PendingSave {
    let Some(cache_path) = delivery.cache_path.clone().filter(|_| !writes.none()) else {
        return PendingSave::none();
    };
    // Entered here, on the caller's thread, rather than by the writers: a run that
    // returns with a pending save is saving from the caller's point of view from this
    // moment, and a poller sees the phase without waiting for a thread to be scheduled.
    if let Some(progress) = progress {
        progress.enter(ProgressPhase::Saving);
    }

    // The index is read-only from here, so the writer and the caller's rendering are two
    // readers of one index rather than of two copies. This used to deep-clone — every
    // boxed entry, both stored copies of every name, and every `BTreeMap` — on the
    // caller's thread, before rendering could start, on every cache-writing run.
    // Sharing is what buys the independence a clone was buying; a run with nothing to
    // write still returns above rather than reaching this point.
    let snapshot_source = std::sync::Arc::clone(index);
    let mut workers = Vec::with_capacity(2);
    if writes.metadata {
        let metadata_source = std::sync::Arc::clone(&snapshot_source);
        let metadata_path = cache_path.clone();
        if let Ok(worker) =
            std::thread::Builder::new().name("fdu-snapshot".to_string()).spawn(move || {
                let _counter_guard = counters::thread_flush_guard();
                let saved = snapshot::save(&metadata_source, &metadata_path);
                // The caller joins this thread; a writer holding the last reference would
                // otherwise make that join wait for the whole index to be freed.
                release_index(metadata_source);
                saved
            })
        {
            workers.push(("metadata", worker));
        }
    }
    if writes.content {
        let content_path = content::content_cache_path(&cache_path);
        if let Ok(worker) =
            std::thread::Builder::new().name("fdu-content-cache".to_string()).spawn(move || {
                let _counter_guard = counters::thread_flush_guard();
                let saved = content::save_content_cache(&snapshot_source, &content_path);
                release_index(snapshot_source);
                saved
            })
        {
            workers.push(("content", worker));
        }
    }
    // A machine that cannot spawn either thread can still answer; it just answers cold
    // next time.
    PendingSave { workers }
}

/// The application cache directory, resolved independently of a scanned root.
///
/// An explicit destination and `FDU_CACHE_DIR` name this directory exactly.
/// `XDG_CACHE_HOME` and platform locations are bases under which `fdu` lives.
/// A relative path is anchored to the process's current directory before I/O.
pub fn default_cache_dir(explicit: Option<&Path>) -> Result<Option<PathBuf>> {
    resolve_cache_dir(
        explicit,
        nonempty_env("FDU_CACHE_DIR"),
        nonempty_env("XDG_CACHE_HOME"),
        platform_cache_dir(),
    )
}

fn resolve_cache_dir(
    explicit: Option<&Path>,
    override_dir: Option<OsString>,
    xdg: Option<OsString>,
    platform_base: Option<PathBuf>,
) -> Result<Option<PathBuf>> {
    let directory = if let Some(explicit) = explicit {
        if explicit.as_os_str().is_empty() {
            return Err(Error::InvalidValue {
                kind: "cache directory",
                value: String::new(),
                hint: "supply a nonempty directory path".to_string(),
            });
        }
        Some(explicit.to_path_buf())
    } else if let Some(override_dir) = override_dir.filter(|value| !value.is_empty()) {
        Some(PathBuf::from(override_dir))
    } else if let Some(xdg) = xdg.filter(|value| !value.is_empty()) {
        Some(PathBuf::from(xdg).join("fdu"))
    } else {
        platform_base.map(|base| base.join("fdu"))
    };
    directory
        .map(|path| std::path::absolute(&path).map_err(|error| Error::io(&path, error)))
        .transpose()
}

/// The conventional metadata snapshot location for a root in the resolved directory.
///
/// The canonical native root path supplies the stable 16-digit lookup key. The file's
/// own header still proves whether it can answer a request.
pub fn default_cache_path_in(root: &Path, explicit: Option<&Path>) -> Result<Option<PathBuf>> {
    let Some(directory) = default_cache_dir(explicit)? else { return Ok(None) };
    let canonical = root.canonicalize().map_err(|error| Error::io(root, error))?;
    let mut hash = 0xcbf2_9ce4_8422_2325_u64;
    for byte in canonical.as_os_str().as_encoded_bytes() {
        hash ^= u64::from(*byte);
        hash = hash.wrapping_mul(0x1000_0000_01b3);
    }
    Ok(Some(CachePaths::for_root_hash(&directory, hash).metadata))
}

/// Conventional metadata snapshot path for callers without an explicit destination.
pub fn default_cache_path(root: &Path) -> Option<PathBuf> {
    default_cache_path_in(root, None).ok().flatten()
}

fn nonempty_env(name: &str) -> Option<OsString> {
    std::env::var_os(name).filter(|value| !value.is_empty())
}

#[cfg(target_os = "windows")]
fn platform_cache_dir() -> Option<PathBuf> {
    windows_cache_dir(
        nonempty_env("LOCALAPPDATA"),
        nonempty_env("USERPROFILE"),
        nonempty_env("HOME"),
    )
}

#[cfg(target_os = "windows")]
fn windows_cache_dir(
    local_app_data: Option<OsString>,
    user_profile: Option<OsString>,
    home: Option<OsString>,
) -> Option<PathBuf> {
    local_app_data
        .map(PathBuf::from)
        .or_else(|| user_profile.map(|path| PathBuf::from(path).join("AppData").join("Local")))
        .or_else(|| home.map(|path| PathBuf::from(path).join(".cache")))
}

#[cfg(target_os = "macos")]
fn platform_cache_dir() -> Option<PathBuf> {
    Some(PathBuf::from(nonempty_env("HOME")?).join(".cache"))
}

#[cfg(not(any(target_os = "windows", target_os = "macos")))]
fn platform_cache_dir() -> Option<PathBuf> {
    Some(PathBuf::from(nonempty_env("HOME")?).join(".cache"))
}

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

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

    /// `fixture`, answered from its snapshot alone.
    fn stale(fixture: OpenFixture) -> OpenFixture {
        OpenFixture { stale_ok: true, ..fixture }
    }

    fn controls_config(
        policy: CachePolicy,
        snapshot_path: PathBuf,
        read_controls: bool,
    ) -> OpenFixture {
        OpenFixture {
            scan: ScanConfig { read_controls, ..ScanConfig::default() },
            cache_path: Some(snapshot_path),
            policy,
            ..OpenFixture::default()
        }
    }

    fn seed_controls_snapshot(root: &Path, snapshot_path: PathBuf) {
        let seed = controls_config(CachePolicy::Auto, snapshot_path, true);
        let (index, report) = open_fixture(root, &seed).expect("seed controls-on snapshot");
        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert!(
            !index.controls().expect("control state observed").is_empty(),
            "the fixture must retain a control source"
        );
    }

    /// Releasing one reference never frees an index another holder can still read: a
    /// caller and a snapshot writer let go in either order, and only the last does any
    /// work.
    #[test]
    fn releasing_a_shared_index_leaves_the_other_holder_reading_it() {
        let root = tempfile::tempdir().expect("tempdir");
        write_file(&root.path().join("kept.txt"), b"kept");
        let (index, _) = open_fixture(root.path(), &OpenFixture::default()).expect("open");
        let shared = std::sync::Arc::new(index);
        let writer = std::sync::Arc::clone(&shared);

        release_index(shared);

        assert_eq!(std::sync::Arc::strong_count(&writer), 1);
        assert_eq!(writer.total().files, 1);
        release_index(writer);
    }

    /// A default `open` observes control state, so its index answers ignore questions
    /// exactly. A request that turns observation off reads no control file, so a control
    /// line no index could retain does not end it, and its index says it cannot classify
    /// ignored entries rather than calling every entry unignored.
    #[test]
    fn a_default_open_answers_ignore_questions_and_an_opt_out_refuses_them() {
        let root = tempfile::tempdir().expect("tempdir");
        write_file(&root.path().join(".gitignore"), b"*.log\n");
        write_file(&root.path().join("debug.log"), b"ignored");
        write_file(&root.path().join("keep.rs"), b"kept");
        let uncached = OpenFixture { policy: CachePolicy::Off, ..OpenFixture::default() };
        let opted_out = OpenFixture {
            scan: ScanConfig { read_controls: false, ..ScanConfig::default() },
            ..uncached.clone()
        };

        let (index, report) = open_fixture(root.path(), &uncached).expect("default open");
        assert!(report.is_complete(), "{:?}", report.errors());
        assert!(index.observes_controls());
        assert_eq!(index.is_ignored(Path::new("debug.log")).ok(), Some(Some(true)));
        assert_eq!(index.is_ignored(Path::new("keep.rs")).ok(), Some(Some(false)));
        assert_eq!(index.is_ignored(Path::new("absent")).ok(), Some(None));
        assert!(
            index
                .controls()
                .is_ok_and(|controls| controls.source_is(Path::new(".gitignore"), b"*.log\n"))
        );
        assert_eq!(index.partition_total().expect("control state observed").unignored.files, 2);

        let mut oversized = vec![b'x'; crate::control::DEFAULT_CONTROL_LINE_LIMIT + 1];
        oversized.extend_from_slice(b"\n*.log\n");
        write_file(&root.path().join(".gitignore"), &oversized);
        let (index, report) =
            open_fixture(root.path(), &uncached).expect("a refused control ends nothing");
        assert!(report.is_complete(), "{:?}", report.errors());
        let crate::control::ControlCoverage::Observed(coverage) = index.control_coverage() else {
            panic!("a default open reads the control file the opt-out skips");
        };
        assert_eq!((coverage.applied, coverage.refused), (0, 1));

        let (index, report) = open_fixture(root.path(), &opted_out)
            .expect("an opted-out open reads no control line, however long");
        assert!(report.is_complete(), "{:?}", report.errors());
        assert!(!index.observes_controls());
        for path in ["debug.log", "keep.rs", "absent"] {
            assert!(
                matches!(index.is_ignored(Path::new(path)), Err(Error::ControlStateNotObserved)),
                "{path} must not be called unignored by an index that read no rule"
            );
        }
        assert!(matches!(index.controls(), Err(Error::ControlStateNotObserved)));
        assert!(matches!(index.partition_total(), Err(Error::ControlStateNotObserved)));
        assert_eq!(index.total().files, 3);
    }

    /// Lifting a control limit scans cold once, and the snapshot it writes then serves
    /// those limits warm, with the coverage it recorded; the other limits' request misses.
    #[test]
    fn a_snapshot_serves_only_the_control_limits_it_was_taken_under() {
        let root = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        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(&root.path().join(".gitignore"), &long_line);
        write_file(&root.path().join("debug.log"), b"ignored");
        let default = controls_config(CachePolicy::Auto, snapshot_path.clone(), true);
        let lifted_limits = crate::control::ControlLimits {
            line_limit: None,
            ..crate::control::ControlLimits::default()
        };
        let lifted = OpenFixture {
            scan: ScanConfig { control_limits: lifted_limits, ..default.scan.clone() },
            ..controls_config(CachePolicy::Auto, snapshot_path, true)
        };
        let refused = |index: &Index| match index.control_coverage() {
            crate::control::ControlCoverage::Observed(coverage) => coverage.refused,
            crate::control::ControlCoverage::NotObserved => panic!("observed"),
        };

        let (index, report) = open_fixture(root.path(), &default).expect("default limits");
        assert_eq!((report.path_taken, refused(&index)), (OpenPath::ColdScan, 1));
        let (index, report) = open_fixture(root.path(), &default).expect("default limits again");
        assert_eq!((report.path_taken, refused(&index)), (OpenPath::WarmRevalidate, 1));

        let (index, report) = open_fixture(root.path(), &lifted).expect("lifted line limit");
        assert_eq!((report.path_taken, refused(&index)), (OpenPath::ColdScan, 0));
        assert_eq!(index.is_ignored(Path::new("debug.log")).ok(), Some(Some(true)));
        let (index, report) = open_fixture(root.path(), &lifted).expect("lifted line limit again");
        assert_eq!((report.path_taken, refused(&index)), (OpenPath::WarmRevalidate, 0));

        let (_, report) = open_fixture(root.path(), &default).expect("back to the default");
        assert_eq!(report.path_taken, OpenPath::ColdScan);
    }

    #[test]
    fn controls_on_snapshot_projects_to_controls_off_auto_open_without_replacing_it() {
        let root = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&root.path().join(".gitignore"), b"ignored.log\n");
        write_file(&root.path().join("ignored.log"), b"ignored");
        seed_controls_snapshot(root.path(), snapshot_path.clone());
        let stronger = fs::read(&snapshot_path).expect("stronger snapshot");
        write_file(&root.path().join("new.txt"), b"new");

        let controls_off = controls_config(CachePolicy::Auto, snapshot_path.clone(), false);
        let (index, report) =
            open_fixture(root.path(), &controls_off).expect("projected warm open");

        assert_eq!(report.path_taken, OpenPath::WarmRevalidate);
        assert!(report.projected);
        assert_eq!(index.scope(), controls_off.scan.scope());
        assert!(matches!(index.controls(), Err(Error::ControlStateNotObserved)));
        assert!(matches!(index.path_state(Path::new("new.txt")), PathState::Present { .. }));
        assert_eq!(fs::read(snapshot_path).expect("snapshot retained"), stronger);
    }

    /// A cache-only open refused for the control limits names them and what recovers.
    ///
    /// Both scopes observe control state and differ only in their ignore-rules identity,
    /// which is a hash: without naming the limits, the caller is told "a different scan
    /// scope" about a request whose only difference is a limit they chose.
    #[test]
    fn a_cache_only_open_after_a_limit_change_names_the_limits_that_differ() {
        let root = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&root.path().join(".gitignore"), b"ignored.log\n");
        write_file(&root.path().join("ignored.log"), b"ignored");
        seed_controls_snapshot(root.path(), snapshot_path.clone());

        let lifted = crate::control::ControlLimits {
            line_limit: None,
            ..crate::control::ControlLimits::default()
        };
        let mut wanted = stale(controls_config(CachePolicy::Auto, snapshot_path, true));
        wanted.scan.control_limits = lifted;
        let Err(Error::Snapshot(message)) = open_fixture(root.path(), &wanted) else {
            panic!("a snapshot taken under other limits must not serve a cache-only open");
        };
        assert!(
            message.contains("line limit 16 KiB, where this request asks for all"),
            "names the limit that differs and both values: {message}"
        );
        assert!(!message.contains("budget"), "the budget is unchanged: {message}");
        assert!(!message.contains("different scan scope"), "says which scope differs: {message}");
        assert!(
            message.contains("repeat the request with the snapshot's limits"),
            "names the remedy: {message}"
        );
        assert!(message.contains("verified answer"), "names the other remedy: {message}");
    }

    #[test]
    fn controls_on_snapshot_projects_to_controls_off_cache_only_open() {
        let root = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&root.path().join(".gitignore"), b"ignored.log\n");
        write_file(&root.path().join("ignored.log"), b"ignored");
        seed_controls_snapshot(root.path(), snapshot_path.clone());

        let controls_off = stale(controls_config(CachePolicy::Auto, snapshot_path, false));
        let (index, report) =
            open_fixture(root.path(), &controls_off).expect("projected cache-only open");
        assert_eq!(report.path_taken, OpenPath::CacheOnly);
        assert!(report.projected);
        assert_eq!(index.scope(), controls_off.scan.scope());
        assert!(matches!(index.controls(), Err(Error::ControlStateNotObserved)));
    }

    #[test]
    fn cache_only_names_foreign_type_rules_after_validating_the_snapshot_and_root() {
        let root = tempfile::tempdir().expect("root");
        let other_root = tempfile::tempdir().expect("other root");
        let cache = tempfile::tempdir().expect("cache");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&root.path().join("main.rs"), b"fn main() {}\n");
        let custom_types = std::sync::Arc::new(
            classify::TypeRegistry::from_manifest(
                "[[kind]]\nid = \"notes\"\nfamily = \"prose\"\nextensions = [\"rs\"]\n",
            )
            .expect("custom rules"),
        );
        let custom = OpenFixture {
            scan: ScanConfig::default().with_types(custom_types.clone()),
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(root.path(), &custom).expect("write custom snapshot");

        let default_only = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            stale_ok: true,
            ..OpenFixture::default()
        };
        let Err(Error::Snapshot(message)) = open_fixture(root.path(), &default_only) else {
            panic!("foreign type rules must not serve cache-only");
        };
        assert!(message.contains("different file type rules"), "{message}");
        assert!(
            message.contains("type registry") && message.contains("verified answer"),
            "{message}"
        );
        assert!(
            snapshot::load(&snapshot_path).expect("direct default-registry load").is_none(),
            "a direct load cannot attach the compiled registry to foreign rules"
        );
        let (_, report) =
            open_fixture(root.path(), &OpenFixture { stale_ok: true, ..custom.clone() })
                .expect("matching registry can use the snapshot");
        assert_eq!(report.path_taken, OpenPath::CacheOnly);

        let Err(Error::Snapshot(message)) = open_fixture(other_root.path(), &default_only) else {
            panic!("a foreign root and rules must not serve cache-only");
        };
        assert!(message.contains("different root"), "root refusal takes precedence: {message}");
        assert!(!message.contains("type rules"), "wrong-root identity must not leak: {message}");

        let mut corrupt = fs::read(&snapshot_path).expect("snapshot bytes");
        let middle = corrupt.len() / 2;
        corrupt[middle] ^= 1;
        fs::write(&snapshot_path, corrupt).expect("corrupt payload without updating checksum");
        let Err(Error::Snapshot(message)) = open_fixture(root.path(), &default_only) else {
            panic!("corrupt snapshot must not yield a type-rules refusal");
        };
        assert!(!message.contains("type rules"), "corruption is an absent snapshot: {message}");
    }

    /// Supplied rules reach the answer, and invalidate a snapshot taken under others.
    ///
    /// The end-to-end property behind the registry: a consumer whose taxonomy differs
    /// from this repository's classifies its own way without rebuilding the crate, and a
    /// snapshot written under one taxonomy is never served under another. The second half
    /// is the one that would fail silently -- the entry counts and byte totals are
    /// identical either way, so a stale snapshot looks entirely correct.
    #[test]
    fn supplied_type_rules_change_the_answer_and_invalidate_the_snapshot() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("main.rs"), b"fn main() {}\n");

        let default_config = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            analysis: content::AnalysisRequest {
                profile: content::AnalysisSet::NONE.with_lines(),
                ..content::AnalysisRequest::default()
            },
            ..OpenFixture::default()
        };
        let (index, _) = open_fixture(dir.path(), &default_config).expect("default open");
        assert_eq!(index.classify(Path::new("main.rs")).file_type.as_str(), "rust");
        drop(index);

        let mine = std::sync::Arc::new(
            classify::TypeRegistry::from_manifest(
                "[[kind]]\nid = \"notes\"\nfamily = \"prose\"\nextensions = [\"rs\"]\n",
            )
            .expect("a minimal manifest"),
        );
        let custom_config = OpenFixture {
            scan: scan::ScanConfig::default().with_types(mine.clone()),
            ..default_config.clone()
        };

        assert_ne!(
            custom_config.scan.scope(),
            default_config.scan.scope(),
            "different rules are a different scan scope"
        );

        let (index, report) = open_fixture(dir.path(), &custom_config).expect("custom open");
        assert_eq!(
            report.path_taken,
            OpenPath::ColdScan,
            "the snapshot was written under other rules and must not be reused"
        );
        assert_eq!(index.classify(Path::new("main.rs")).file_type.as_str(), "notes");
        assert_eq!(index.types().fingerprint(), mine.fingerprint());
        let content = index
            .content()
            .and_then(|content| content.file(Path::new("main.rs")))
            .expect("custom analysis record");
        assert_eq!(content.detection.file_type.as_str(), "notes");
        assert_eq!(
            index
                .content()
                .and_then(content::ContentIndex::provenance)
                .expect("content provenance")
                .type_rules_fingerprint,
            mine.fingerprint()
        );

        // And the snapshot the custom run wrote is reusable by a run under the same rules.
        let (_, report) = open_fixture(dir.path(), &custom_config).expect("second custom open");
        assert_eq!(report.path_taken, OpenPath::WarmRevalidate, "same rules, same snapshot");
        assert_eq!(report.content_cache.hits, 1, "the matching sidecar is reusable");
        assert_eq!(report.analysis.expect("analysis report").candidates, 0);

        assert!(
            snapshot::load(&snapshot_path).expect("default-registry load").is_none(),
            "a direct default-registry load must reject a custom-registry snapshot"
        );
        let loaded = snapshot::load_with_types(&snapshot_path, mine)
            .expect("custom-registry load")
            .expect("the matching custom registry makes the snapshot usable");
        assert_eq!(loaded.classify(Path::new("main.rs")).file_type.as_str(), "notes");
    }

    /// The behaviour table from the design, asserted rather than described.
    #[test]
    fn each_cache_policy_reads_scans_and_writes_as_documented() {
        // An `open` is its own later reader, so `Auto` writes here; the one-shot half of
        // the table is `execution`'s `auto_persists_where_a_later_request_reads_what_it_stores`.
        for (policy, expect_write) in
            [(CachePolicy::Auto, true), (CachePolicy::On, true), (CachePolicy::Off, false)]
        {
            let dir = tempfile::tempdir().expect("tempdir");
            let cache = tempfile::tempdir().expect("cache dir");
            let snapshot_path = cache.path().join("snap.fdu");
            write_file(&dir.path().join("a.txt"), b"hello");

            let config = OpenFixture {
                cache_path: Some(snapshot_path.clone()),
                policy,
                ..OpenFixture::default()
            };
            let (index, report) = open_fixture(dir.path(), &config).expect("open");

            assert_eq!(index.total().files, 1, "{policy:?} lost an entry");
            assert_eq!(report.path_taken, OpenPath::ColdScan, "{policy:?} without a snapshot");
            assert_eq!(
                snapshot_path.exists(),
                expect_write,
                "{policy:?} wrote a snapshot: {}",
                snapshot_path.exists()
            );
        }
    }

    /// A verified warm open over an unchanged tree rewrote a byte-identical snapshot on
    /// every run, paying a full index clone, encode, and write to reproduce the file it
    /// had just read.  The bytes are the assertion: if a future change makes an
    /// unchanged reconciliation produce different serialized state, this fails loudly
    /// rather than letting the skip silently drop it.
    #[test]
    fn an_unchanged_warm_open_leaves_the_snapshot_alone_and_a_changed_one_rewrites_it() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");
        write_file(&dir.path().join("sub/b.txt"), b"world");

        let auto = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &auto).expect("seed");
        let seeded = fs::read(&snapshot_path).expect("seeded snapshot");

        let (index, report) = open_fixture(dir.path(), &auto).expect("warm open");
        assert_eq!(report.path_taken, OpenPath::WarmRevalidate);
        assert_eq!(index.total().files, 2);
        assert_eq!(
            fs::read(&snapshot_path).expect("snapshot still there"),
            seeded,
            "an unchanged warm open must not rewrite the snapshot"
        );

        write_file(&dir.path().join("sub/c.txt"), b"new file");
        let (index, report) = open_fixture(dir.path(), &auto).expect("warm open after a change");
        assert_eq!(report.path_taken, OpenPath::WarmRevalidate);
        assert_eq!(index.total().files, 3);
        let after_add = fs::read(&snapshot_path).expect("rewritten snapshot");
        assert_ne!(after_add, seeded, "a warm open that found a new file must persist it");

        fs::remove_file(dir.path().join("sub/c.txt")).expect("remove");
        let (index, _) = open_fixture(dir.path(), &auto).expect("warm open after a removal");
        assert_eq!(index.total().files, 2);
        assert_ne!(
            fs::read(&snapshot_path).expect("rewritten snapshot"),
            after_add,
            "a warm open that found a removal must persist it"
        );

        // The skip must leave a snapshot a later cache-only open can still serve.
        let cache_only = OpenFixture { stale_ok: true, ..auto };
        let (restored, _) = open_fixture(dir.path(), &cache_only).expect("cache-only open");
        assert_eq!(restored.total().files, 2);
    }

    #[test]
    fn cache_only_answers_from_the_snapshot_without_touching_the_tree() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");

        let auto = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &auto).expect("seed");

        // Change the tree after the snapshot was taken. A cache-only answer must report
        // what it has, not what is there now — and its freshness must say so.
        write_file(&dir.path().join("b.txt"), b"new file");

        let only = OpenFixture { stale_ok: true, ..auto };
        let (index, report) = open_fixture(dir.path(), &only).expect("cache-only open");
        assert_eq!(report.path_taken, OpenPath::CacheOnly);
        assert_eq!(index.total().files, 1, "the new file must not appear");
        assert_ne!(index.freshness(), Freshness::Fresh, "a stale answer must not claim currency");
    }

    #[test]
    fn cache_only_fails_closed_when_no_snapshot_is_usable() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&dir.path().join("a.txt"), b"hello");

        // Guessing a scan here would make the fast path unpredictable: sometimes instant,
        // sometimes a full walk, with nothing in the output to say which happened.
        let only = OpenFixture {
            cache_path: Some(cache.path().join("absent.fdu")),
            stale_ok: true,
            ..OpenFixture::default()
        };
        let Err(Error::Snapshot(message)) = open_fixture(dir.path(), &only) else {
            panic!("a cache-only open with no snapshot must fail");
        };
        // A diagnostic names its remedy: a stale answer is the one delivery that cannot
        // recover, so the message says what leaves a snapshot and what scans instead.
        assert!(message.contains("with the `on` cache policy"), "names the remedy: {message}");
        assert!(message.contains("verified answer"), "names the other remedy: {message}");
    }

    #[test]
    fn content_sidecar_skips_unchanged_reads_and_serves_cache_only() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("notes.md"), b"one two\n");
        let analysis = content::AnalysisRequest {
            profile: content::AnalysisSet::NONE.with_lines(),
            ..content::AnalysisRequest::default()
        };
        let auto = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            analysis,
            ..OpenFixture::default()
        };

        let (first, first_report) = open_fixture(dir.path(), &auto).expect("cold analyzed open");
        assert_eq!(first_report.analysis.expect("analysis").lines.analyzed, 1);
        assert_eq!(
            first.content_rollup(Path::new("")).expect("content").total.lines.metrics.raw_words,
            2
        );
        assert!(content::content_cache_path(&snapshot_path).exists());

        let (_, warm_report) = open_fixture(dir.path(), &auto).expect("warm analyzed open");
        assert_eq!(warm_report.content_cache.hits, 1);
        assert_eq!(warm_report.content_cache.bytes, 8);
        assert_eq!(warm_report.analysis.expect("analysis").candidates, 0);

        fs::remove_file(dir.path().join("notes.md")).expect("remove source");
        let only = OpenFixture { stale_ok: true, ..auto };
        let (cached, cached_report) = open_fixture(dir.path(), &only).expect("cache-only content");
        assert_eq!(cached_report.content_cache.hits, 1);
        assert_eq!(cached_report.content_cache.bytes, 8);
        assert_eq!(
            cached.content_rollup(Path::new("")).expect("content").total.lines.metrics.raw_words,
            2
        );
    }

    #[test]
    fn narrowed_population_snapshot_and_sidecar_reuse_only_their_scope() {
        let dir = tempfile::tempdir().expect("tree");
        let cache = tempfile::tempdir().expect("cache");
        let snapshot_path = cache.path().join("scoped.metadata.bin");
        write_file(&dir.path().join(".gitignore"), b"*.log\n");
        write_file(&dir.path().join("keep.rs"), b"kept\n");
        write_file(&dir.path().join("debug.log"), b"ignored\n");
        let analysis = content::AnalysisRequest {
            profile: content::AnalysisSet::NONE.with_lines(),
            ..content::AnalysisRequest::default()
        };
        let excluded = OpenFixture {
            stale_ok: false,
            scan: ScanConfig {
                population: query::IgnoredEntries::Exclude,
                ..ScanConfig::default()
            },
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            analysis,
        };
        let (cold, report) = open_fixture(dir.path(), &excluded).expect("cold exclude");
        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert!(cold.lookup(Path::new("keep.rs")).is_some());
        assert!(cold.lookup(Path::new("debug.log")).is_none());
        assert!(content::content_cache_path(&snapshot_path).exists());

        let only_cache = OpenFixture { stale_ok: true, ..excluded.clone() };
        let (restored, report) = open_fixture(dir.path(), &only_cache).expect("cache-only exclude");
        assert_eq!(report.path_taken, OpenPath::CacheOnly);
        assert!(report.content_cache.hits > 0);
        assert!(restored.lookup(Path::new("keep.rs")).is_some());
        assert!(restored.lookup(Path::new("debug.log")).is_none());

        let opposite = OpenFixture {
            scan: ScanConfig { population: query::IgnoredEntries::Only, ..ScanConfig::default() },
            ..only_cache
        };
        assert!(matches!(open_fixture(dir.path(), &opposite), Err(Error::Snapshot(_))));
    }

    #[test]
    fn cached_coverage_exclusions_remain_visible_without_making_the_run_partial() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("invalid.txt"), b"valid prefix\xff");
        let auto = OpenFixture {
            cache_path: Some(snapshot_path),
            policy: CachePolicy::Auto,
            analysis: content::AnalysisRequest {
                profile: content::AnalysisSet::NONE.with_lines(),
                ..content::AnalysisRequest::default()
            },
            ..OpenFixture::default()
        };

        let (_, cold_report) = open_fixture(dir.path(), &auto).expect("cold analyzed open");
        assert!(cold_report.is_complete());
        assert_eq!(cold_report.analysis.expect("analysis").lines.invalid_utf8, 1);
        assert!(cold_report.error_messages().is_empty());

        let (_, warm_report) = open_fixture(dir.path(), &auto).expect("warm analyzed open");
        assert_eq!(warm_report.content_cache.hits, 1);
        assert_eq!(warm_report.content_cache.coverage_exclusions, 1);
        assert_eq!(warm_report.analysis.expect("analysis").candidates, 0);
        assert!(warm_report.is_complete());
        assert!(warm_report.error_messages().is_empty());

        let only = OpenFixture { stale_ok: true, ..auto };
        let (_, cached_report) = open_fixture(dir.path(), &only).expect("cache-only analyzed open");
        assert_eq!(cached_report.content_cache.coverage_exclusions, 1);
        assert!(cached_report.is_complete());
        assert!(cached_report.error_messages().is_empty());
    }

    #[test]
    fn cache_only_analysis_fails_closed_without_its_sidecar() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("notes.md"), b"one two\n");
        let metadata_only = OpenFixture {
            cache_path: Some(snapshot_path),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &metadata_only).expect("seed metadata");

        let only = OpenFixture {
            stale_ok: true,
            analysis: content::AnalysisRequest {
                profile: content::AnalysisSet::NONE.with_lines(),
                ..content::AnalysisRequest::default()
            },
            ..metadata_only
        };
        assert!(matches!(open_fixture(dir.path(), &only), Err(Error::Snapshot(_))));

        // A sidecar of another analyzer set is not this request's, whether it is wider or
        // narrower: cache-only fails closed rather than answering with the stored set.
        let with_set = |policy, profile| OpenFixture {
            policy,
            stale_ok: false,
            analysis: content::AnalysisRequest { profile, ..content::AnalysisRequest::default() },
            ..only.clone()
        };
        let lines = content::AnalysisSet::NONE.with_lines();
        for (stored, wanted) in
            [(content::AnalysisSet::ALL, lines), (lines, content::AnalysisSet::ALL)]
        {
            open_fixture(dir.path(), &with_set(CachePolicy::Auto, stored))
                .expect("write a sidecar");
            let refused = open_fixture(dir.path(), &stale(with_set(CachePolicy::Auto, wanted)));
            assert!(
                matches!(refused, Err(Error::Snapshot(_))),
                "a {stored:?} sidecar must not serve a cache-only {wanted:?} request"
            );
        }

        open_fixture(dir.path(), &with_set(CachePolicy::Auto, lines))
            .expect("write the lines sidecar");
        let (cached, report) = open_fixture(dir.path(), &only).expect("restore the analyzed state");
        assert!(report.content_cache.usable);
        assert_eq!(report.content_cache.hits, 1, "one record per candidate is complete");
        assert_eq!(
            report.content_cache.candidates, report.content_cache.hits,
            "completeness uses the count restore already walked"
        );
        assert_eq!(cached.content_set(), lines);
    }

    #[test]
    fn cache_only_analysis_fails_closed_when_the_sidecar_is_incomplete() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("notes.md"), b"one two\n");
        let analysis = content::AnalysisRequest {
            profile: content::AnalysisSet::NONE.with_lines(),
            ..content::AnalysisRequest::default()
        };
        let auto = OpenFixture {
            cache_path: Some(snapshot_path),
            policy: CachePolicy::Auto,
            analysis,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &auto).expect("seed one analyzed file");

        // A later metadata-only pass widens the snapshot without rewriting the sidecar,
        // so cache-only analysis must refuse rather than report a partial content answer.
        write_file(&dir.path().join("extra.txt"), b"three\n");
        let metadata_only =
            OpenFixture { analysis: content::AnalysisRequest::default(), ..auto.clone() };
        open_fixture(dir.path(), &metadata_only).expect("widen the snapshot");

        let only = OpenFixture { stale_ok: true, ..auto };
        assert!(
            matches!(open_fixture(dir.path(), &only), Err(Error::Snapshot(_))),
            "a one-record sidecar must not serve a two-file cache-only analysis request"
        );
    }

    #[cfg(unix)]
    #[test]
    fn cache_only_serves_checksummed_native_non_utf8_names() {
        use std::ffi::OsString;
        use std::os::unix::ffi::OsStringExt;

        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        let root = dir.path().canonicalize().expect("canonical root");
        let native = PathBuf::from(OsString::from_vec(vec![b'n', 0x80]));
        let mut index = Index::new(&root);
        index.apply_ok(&Observation::new(vec![
            Op::Upsert {
                path: PathBuf::from("ok.txt"),
                kind: EntryKind::File,
                attrs: Attrs {
                    size: 1,
                    allocated: 512,
                    mtime_ns: 1,
                    ctime_ns: 1,
                    inode: 1,
                    dev: 1,
                },
            },
            Op::Upsert {
                path: native.clone(),
                kind: EntryKind::File,
                attrs: Attrs {
                    size: 2,
                    allocated: 512,
                    mtime_ns: 2,
                    ctime_ns: 2,
                    inode: 2,
                    dev: 1,
                },
            },
        ]));
        snapshot::save(&index, &snapshot_path).expect("save native names");

        let only = OpenFixture {
            cache_path: Some(snapshot_path),
            stale_ok: true,
            ..OpenFixture::default()
        };
        let (cached, report) = open_fixture(&root, &only).expect("cache-only native name");
        assert_eq!(cached.total().files, 2);
        assert!(cached.lookup(&native).is_some());
        assert!(report.is_complete());
    }

    #[test]
    fn cache_only_empty_analysis_still_requires_a_usable_sidecar() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        let metadata_only = OpenFixture {
            cache_path: Some(snapshot_path),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &metadata_only).expect("seed empty metadata");

        let only = OpenFixture {
            stale_ok: true,
            analysis: content::AnalysisRequest {
                profile: content::AnalysisSet::NONE.with_lines(),
                ..content::AnalysisRequest::default()
            },
            ..metadata_only
        };
        assert!(matches!(open_fixture(dir.path(), &only), Err(Error::Snapshot(_))));
    }

    #[test]
    fn a_snapshot_for_another_root_is_treated_as_absent() {
        let one = tempfile::tempdir().expect("tempdir");
        let two = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&one.path().join("a.txt"), b"hello");
        write_file(&two.path().join("b.txt"), b"other tree");

        let config = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(one.path(), &config).expect("seed from the first root");

        // Reading another tree's snapshot would be worse than a cache miss.
        let (index, report) = open_fixture(two.path(), &config).expect("second root");
        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert_eq!(index.total().files, 1);
        assert_eq!(index.root_path(), two.path().canonicalize().expect("canonical").as_path());
    }

    #[test]
    fn open_without_a_cache_always_scans_cold() {
        let dir = tempfile::tempdir().expect("tempdir");
        write_file(&dir.path().join("a.txt"), b"hello");

        let (index, report) = open_fixture(dir.path(), &OpenFixture::default()).expect("open");
        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert_eq!(index.total().files, 1);
    }

    #[test]
    fn second_open_takes_the_warm_path_and_stays_correct() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&dir.path().join("a.txt"), b"hello");
        write_file(&dir.path().join("src/main.rs"), b"fn main() {}");

        let config = OpenFixture {
            cache_path: Some(cache.path().join("snap.fdu")),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };

        let (first, first_report) = open_fixture(dir.path(), &config).expect("cold open");
        assert_eq!(first_report.path_taken, OpenPath::ColdScan);
        assert_eq!(first.total().files, 2);

        // Change the tree between opens: the warm path must notice.
        write_file(&dir.path().join("added.md"), b"new");
        fs::remove_file(dir.path().join("a.txt")).expect("remove");

        let (second, second_report) = open_fixture(dir.path(), &config).expect("warm open");
        assert_eq!(second_report.path_taken, OpenPath::WarmRevalidate);
        assert_eq!(second.total().files, 2);
        assert!(second.lookup(Path::new("added.md")).is_some());
        assert!(second.lookup(Path::new("a.txt")).is_none());
    }

    #[test]
    fn a_snapshot_from_another_root_is_ignored() {
        let a = tempfile::tempdir().expect("tempdir a");
        let b = tempfile::tempdir().expect("tempdir b");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&a.path().join("only-in-a.txt"), b"x");
        write_file(&b.path().join("only-in-b.txt"), b"y");

        let cache_path = cache.path().join("snap.fdu");
        let config = OpenFixture {
            cache_path: Some(cache_path),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };

        open_fixture(a.path(), &config).expect("open a");
        let (index, report) = open_fixture(b.path(), &config).expect("open b");

        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert!(index.lookup(Path::new("only-in-b.txt")).is_some());
        assert!(index.lookup(Path::new("only-in-a.txt")).is_none());
    }

    #[test]
    fn snapshot_scope_mismatch_forces_a_cold_scan() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&dir.path().join("top.txt"), b"top");
        write_file(&dir.path().join("deep/nested.txt"), b"nested");

        let cache_path = cache.path().join("snap.fdu");
        let full = OpenFixture {
            cache_path: Some(cache_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &full).expect("full open");

        let shallow = OpenFixture {
            scan: ScanConfig { max_depth: Some(1), ..ScanConfig::default() },
            cache_path: Some(cache_path),
            ..OpenFixture::default()
        };
        let (index, report) = open_fixture(dir.path(), &shallow).expect("shallow open");

        assert_eq!(report.path_taken, OpenPath::ColdScan);
        assert!(index.lookup(Path::new("deep")).is_some());
        assert!(index.lookup(Path::new("deep/nested.txt")).is_none());
    }

    #[test]
    fn admission_scope_mismatch_cannot_reinterpret_a_snapshot() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&dir.path().join(".hidden"), b"hidden");
        let cache_path = cache.path().join("snap.fdu");
        let seed = OpenFixture {
            cache_path: Some(cache_path.clone()),
            policy: CachePolicy::Auto,
            ..OpenFixture::default()
        };
        open_fixture(dir.path(), &seed).expect("seed snapshot");

        let changed_scopes = [
            ScanConfig {
                hidden: Some(std::sync::Arc::new(
                    HiddenPolicy::prune_hidden::<[&str; 0], &str>([]),
                )),
                ..ScanConfig::default()
            },
            ScanConfig { exclude_special: true, ..ScanConfig::default() },
        ];
        for scan in changed_scopes {
            let only = OpenFixture {
                policy: CachePolicy::Auto,
                scan,
                cache_path: Some(cache_path.clone()),
                stale_ok: true,
                analysis: content::AnalysisRequest::default(),
            };
            assert!(matches!(open_fixture(dir.path(), &only), Err(Error::Snapshot(_))));
        }
    }

    #[test]
    fn operational_batch_size_does_not_invalidate_a_snapshot() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        write_file(&dir.path().join("a.txt"), b"a");
        let cache_path = cache.path().join("snap.fdu");

        let first = OpenFixture {
            stale_ok: false,
            scan: ScanConfig { batch_size: 1, ..ScanConfig::default() },
            cache_path: Some(cache_path.clone()),
            policy: CachePolicy::Auto,
            analysis: content::AnalysisRequest::default(),
        };
        open_fixture(dir.path(), &first).expect("first open");

        let second = OpenFixture {
            scan: ScanConfig { batch_size: 17, ..ScanConfig::default() },
            cache_path: Some(cache_path),
            ..OpenFixture::default()
        };
        let (_, report) = open_fixture(dir.path(), &second).expect("second open");
        assert_eq!(report.path_taken, OpenPath::WarmRevalidate);
    }

    #[test]
    fn cache_paths_differ_per_root() {
        let a = tempfile::tempdir().expect("tempdir a");
        let b = tempfile::tempdir().expect("tempdir b");
        let (Some(pa), Some(pb)) = (default_cache_path(a.path()), default_cache_path(b.path()))
        else {
            return; // No HOME in this environment; nothing to assert.
        };
        assert_ne!(pa, pb);
        assert_eq!(pa, default_cache_path(a.path()).expect("stable"));
    }

    #[test]
    fn cache_destination_precedence_and_pairing() {
        let fixture = tempfile::tempdir().expect("tempdir");
        let explicit = fixture.path().join("explicit");
        let override_dir = fixture.path().join("override");
        let xdg = fixture.path().join("xdg");
        let xdg_app_dir = xdg.join("fdu");
        let native = fixture.path().join("native");
        let resolve = |choice| {
            resolve_cache_dir(
                choice,
                Some(override_dir.as_os_str().to_os_string()),
                Some(xdg.as_os_str().to_os_string()),
                Some(native.clone()),
            )
            .expect("resolve")
            .expect("directory")
        };
        assert_eq!(resolve(Some(&explicit)), explicit);
        assert_eq!(resolve(None), override_dir);
        assert_eq!(
            resolve_cache_dir(None, None, Some(xdg.into_os_string()), Some(native.clone()))
                .expect("resolve"),
            Some(xdg_app_dir)
        );
        assert_eq!(
            resolve_cache_dir(None, None, None, Some(native.clone())).expect("resolve"),
            Some(native.join("fdu"))
        );
        assert!(resolve_cache_dir(Some(Path::new("")), None, None, None).is_err());

        let metadata = default_cache_path_in(fixture.path(), Some(&explicit))
            .expect("path")
            .expect("directory");
        let paths = CachePaths::from_metadata(&metadata);
        assert_eq!(paths.metadata.parent(), Some(explicit.as_path()));
        assert_eq!(
            paths.metadata.file_name().expect("name").to_string_lossy().len(),
            16 + ".metadata.bin".len()
        );
        assert!(
            paths.metadata.file_name().expect("name").to_string_lossy().ends_with(".metadata.bin")
        );
        assert!(
            paths.analysis.file_name().expect("name").to_string_lossy().ends_with(".analysis.bin")
        );
        assert_eq!(paths.analysis.parent(), paths.metadata.parent());
    }

    #[cfg(target_os = "macos")]
    #[test]
    fn macos_native_cache_base_is_under_home_dot_cache() {
        let home = PathBuf::from(nonempty_env("HOME").expect("HOME for native cache"));
        assert_eq!(platform_cache_dir(), Some(home.join(".cache")));
    }

    #[cfg(target_os = "windows")]
    #[test]
    fn windows_cache_discovery_prefers_native_locations() {
        let local = OsString::from(r"C:\Users\tester\AppData\Local");
        let profile = OsString::from(r"D:\Profile");
        let home = OsString::from(r"E:\Home");

        assert_eq!(
            windows_cache_dir(Some(local.clone()), Some(profile.clone()), Some(home.clone())),
            Some(PathBuf::from(local))
        );
        assert_eq!(
            windows_cache_dir(None, Some(profile.clone()), Some(home.clone())),
            Some(PathBuf::from(profile).join("AppData").join("Local"))
        );
        assert_eq!(
            windows_cache_dir(None, None, Some(home.clone())),
            Some(PathBuf::from(home).join(".cache"))
        );
        assert_eq!(windows_cache_dir(None, None, None), None);
    }
}

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

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

    fn config(snapshot_path: &Path, policy: CachePolicy) -> OpenFixture {
        OpenFixture {
            cache_path: Some(snapshot_path.to_path_buf()),
            policy,
            ..OpenFixture::default()
        }
    }

    #[test]
    fn a_pending_save_completes_when_joined() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");

        let (_index, _report, pending) =
            open_fixture_with_pending_save(dir.path(), &config(&snapshot_path, CachePolicy::Auto))
                .expect("open");
        pending.join().expect("save succeeds");
        assert!(snapshot_path.exists(), "a joined save must have landed");
    }

    #[test]
    fn a_dropped_save_still_lands() {
        // Dropping without joining is a caller mistake, not a reason to lose the write:
        // the next run would otherwise pay for a cold scan this one already did.
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");

        {
            let (_index, _report, _pending) = open_fixture_with_pending_save(
                dir.path(),
                &config(&snapshot_path, CachePolicy::Auto),
            )
            .expect("open");
        }
        assert!(snapshot_path.exists());
    }

    #[test]
    #[cfg(unix)]
    fn a_partial_scan_leaves_the_previous_snapshot_alone() {
        // Writing a partial view would serve it as fact on the next run, and the
        // existing complete snapshot is better than that.
        use std::os::unix::fs::PermissionsExt;

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

        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");

        let settings = config(&snapshot_path, CachePolicy::Auto);
        open_fixture(dir.path(), &settings).expect("seed a complete snapshot");
        let complete_len = fs::metadata(&snapshot_path).expect("exists").len();

        let denied = dir.path().join("denied");
        fs::create_dir(&denied).expect("create");
        write_file(&denied.join("hidden.txt"), b"hidden");
        fs::set_permissions(&denied, fs::Permissions::from_mode(0o000)).expect("deny");

        let opened = open_fixture(dir.path(), &settings);
        fs::set_permissions(&denied, fs::Permissions::from_mode(0o700)).expect("restore");
        let (_index, report) = opened.expect("partial open still returns a result");

        assert!(!report.is_complete(), "the scan should be partial");
        assert_eq!(
            fs::metadata(&snapshot_path).expect("still there").len(),
            complete_len,
            "a partial scan must not overwrite a complete snapshot"
        );
    }

    /// Each tier is written by its own rule: a partial pass writes no snapshot, because an
    /// absent entry would change totals, but may write every verified content record beside
    /// a stored snapshot of the same entry identity. It still writes nothing under another
    /// identity, where replacing the sidecar would separate it from the snapshot it names.
    #[test]
    #[cfg(unix)]
    fn a_partial_scan_writes_only_verified_content_for_the_stored_entry_tier() {
        use std::os::unix::fs::PermissionsExt;

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

        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        let sidecar_path = content::content_cache_path(&snapshot_path);
        write_file(&dir.path().join("notes.md"), b"one two\n");
        write_file(&dir.path().join("locked/old.md"), b"three\n");
        let analysis = content::AnalysisRequest {
            profile: content::AnalysisSet::NONE.with_lines(),
            ..content::AnalysisRequest::default()
        };
        let auto = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            analysis,
            ..OpenFixture::default()
        };
        let (_, seeded) = open_fixture(dir.path(), &auto).expect("seed both tiers");
        assert!(seeded.is_complete());
        let snapshot_before = fs::read(&snapshot_path).expect("a snapshot");

        // Change a file the next pass can verify, and lock the directory holding another.
        write_file(&dir.path().join("notes.md"), b"one two three\n");
        let locked = dir.path().join("locked");
        let run = |config: &OpenFixture| {
            fs::set_permissions(&locked, fs::Permissions::from_mode(0o000)).expect("deny");
            let opened = open_fixture(dir.path(), config);
            fs::set_permissions(&locked, fs::Permissions::from_mode(0o700)).expect("restore");
            let (_, report) = opened.expect("a partial open still answers");
            assert!(!report.is_complete(), "the pass should be partial");
        };

        // Under another entry identity: neither tier is written.
        let sidecar_before = fs::read(&sidecar_path).expect("a sidecar");
        let other_scope = OpenFixture {
            scan: ScanConfig { max_depth: Some(8), ..ScanConfig::default() },
            ..auto.clone()
        };
        run(&other_scope);
        assert_eq!(fs::read(&snapshot_path).expect("snapshot"), snapshot_before);
        assert_eq!(
            fs::read(&sidecar_path).expect("sidecar"),
            sidecar_before,
            "a partial run under another entry identity must not evict the paired sidecar"
        );

        // Under the stored snapshot's identity, the snapshot stays whole while the sidecar
        // keeps only content whose entry facts this pass verified.
        run(&auto);
        assert_eq!(
            fs::read(&snapshot_path).expect("snapshot"),
            snapshot_before,
            "a partial scan must not overwrite a complete snapshot"
        );
        assert_ne!(fs::read(&sidecar_path).expect("sidecar"), sidecar_before);
        // The verified changed file is reusable; no record survives under the directory
        // this pass could not list.
        let (mut fresh, _) =
            scan::scan_into_index(dir.path(), &ScanConfig::default()).expect("scan");
        let wanted = fresh.content_identity(analysis.profile);
        let loaded = content::load_content_cache(&mut fresh, &wanted, &sidecar_path).expect("load");
        assert_eq!((loaded.usable, loaded.hits, loaded.stale), (true, 1, 0), "{loaded:?}");
        let content = fresh.content().expect("content");
        assert!(content.file(Path::new("locked/old.md")).is_none());
        assert_eq!(
            content
                .file(Path::new("notes.md"))
                .expect("verified record")
                .lines
                .value()
                .expect("line metrics")
                .raw_words,
            3
        );
    }

    /// A warm partial pass keeps every record whose entry the pass verified, including
    /// unchanged files, and removes records below an unlistable directory.
    #[test]
    #[cfg(unix)]
    fn a_warm_partial_pass_keeps_all_and_only_verified_content_records() {
        use std::os::unix::fs::PermissionsExt;

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

        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        let sidecar_path = content::content_cache_path(&snapshot_path);
        for index in 0..7 {
            write_file(&dir.path().join(format!("f{index}.md")), b"alpha\nbeta\n");
        }
        write_file(&dir.path().join("locked/old.md"), b"three\n");
        let analysis = content::AnalysisRequest {
            profile: content::AnalysisSet::NONE.with_lines(),
            ..content::AnalysisRequest::default()
        };
        let auto = OpenFixture {
            cache_path: Some(snapshot_path.clone()),
            policy: CachePolicy::Auto,
            analysis,
            ..OpenFixture::default()
        };

        let records_in_sidecar = |path: &Path| {
            let (mut fresh, _) = scan::scan_into_index(path, &ScanConfig::default()).expect("scan");
            let wanted = fresh.content_identity(analysis.profile);
            let loaded =
                content::load_content_cache(&mut fresh, &wanted, &sidecar_path).expect("load");
            assert!(loaded.usable, "the sidecar should still be readable");
            loaded.hits + loaded.stale
        };

        let (_, seeded) = open_fixture(dir.path(), &auto).expect("seed both tiers");
        assert!(seeded.is_complete());
        assert_eq!(records_in_sidecar(dir.path()), 8, "one record per seeded file");

        // One file changes, one directory becomes unlistable: the warm pass is partial.
        write_file(&dir.path().join("f0.md"), b"alpha\nbeta\ngamma\n");
        let locked = dir.path().join("locked");
        fs::set_permissions(&locked, fs::Permissions::from_mode(0o000)).expect("deny");
        let opened = open_fixture(dir.path(), &auto);
        fs::set_permissions(&locked, fs::Permissions::from_mode(0o700)).expect("restore");
        let (_, report) = opened.expect("a partial open still answers");
        assert!(!report.is_complete(), "the pass should be partial");

        assert_eq!(
            records_in_sidecar(dir.path()),
            7,
            "all seven verified files survive and the inaccessible record does not"
        );
    }

    #[test]
    fn no_snapshot_is_written_when_the_policy_forbids_it() {
        let dir = tempfile::tempdir().expect("tempdir");
        let cache = tempfile::tempdir().expect("cache dir");
        let snapshot_path = cache.path().join("snap.fdu");
        write_file(&dir.path().join("a.txt"), b"hello");

        let (_index, _report, pending) =
            open_fixture_with_pending_save(dir.path(), &config(&snapshot_path, CachePolicy::Off))
                .expect("open");
        pending.join().expect("nothing to join");
        assert!(!snapshot_path.exists(), "cache off wrote a snapshot");
    }
}