mbx_cache_core/agent/

file_digest.rs

use crate::CacheDigest;
use serde::{Deserialize, Serialize};
use std::io::{self, Read as _};
use std::path::{Path, PathBuf};
#[cfg(target_os = "linux")]
use std::sync::{Mutex, OnceLock};
use std::time::SystemTime;

const DIGEST_BUFFER_BYTES: usize = 64 * 1024;
const TIMESTAMP_MACROS: &[&[u8]] = &[b"__DATE__", b"__TIME__", b"__TIMESTAMP__"];

#[cfg(target_os = "linux")]
const STATX_IDENTITY_MASK: u32 = 0x100 | 0x200 | 0x40 | 0x1000;

/// Stable Linux UAPI layout. `libc` omits statx on older musl headers even
/// though the kernel syscall and wire structure are available there.
#[cfg(target_os = "linux")]
#[repr(C)]
struct LinuxStatxTimestamp {
    seconds: i64,
    nanos: u32,
    reserved: i32,
}

#[cfg(target_os = "linux")]
#[repr(C)]
struct LinuxStatx {
    mask: u32,
    block_size: u32,
    attributes: u64,
    links: u32,
    uid: u32,
    gid: u32,
    mode: u16,
    reserved0: u16,
    inode: u64,
    size: u64,
    blocks: u64,
    attributes_mask: u64,
    accessed: LinuxStatxTimestamp,
    created: LinuxStatxTimestamp,
    changed: LinuxStatxTimestamp,
    modified: LinuxStatxTimestamp,
    rdev_major: u32,
    rdev_minor: u32,
    device_major: u32,
    device_minor: u32,
    mount_id: u64,
    direct_io_memory_alignment: u32,
    direct_io_offset_alignment: u32,
    subvolume: u64,
    atomic_write_unit_min: u32,
    atomic_write_unit_max: u32,
    atomic_write_segments_max: u32,
    direct_io_read_offset_alignment: u32,
    atomic_write_unit_max_opt: u32,
    reserved1: u32,
    reserved2: [u64; 8],
}

#[cfg(target_os = "linux")]
const _: () = assert!(std::mem::size_of::<LinuxStatx>() == 256);

/// The on-disk identity a recorded file digest describes.
///
/// The same trade [`VerifiedBlob`] makes for CAS reads, offered to shims for
/// the files they hash: an overwrite moves the modification time and a
/// truncation changes the length, so a digest recorded against both stands
/// until either does. Where the platform reports a metadata-change time the
/// identity carries that too, and it is the part a writer cannot restore: a
/// rewrite that puts the modification time back still moves the change time.
/// Unix object identity also distinguishes an atomic replacement when both
/// files happen to share one filesystem timestamp tick.
#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct FileIdentity {
    /// Absolute path of the file.
    pub path: PathBuf,
    /// Length of the file in bytes.
    pub len: u64,
    /// Modification time of the file.
    pub modified: SystemTime,
    /// Platform metadata-change token, where one exists.
    pub changed: Option<(i64, i64)>,
    /// Stable object identity used to detect replacement independently of
    /// timestamp resolution.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub object: Option<FileObjectIdentity>,
}

/// The kernel identity of one file object on one mounted filesystem.
#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct FileObjectIdentity {
    /// Device identity component. Linux NFS uses the major number reported by
    /// `statx`; other Unix filesystems store the high half of `st_dev`.
    pub device_major: u32,
    /// Device identity component. Linux NFS uses the minor number reported by
    /// `statx`; other Unix filesystems store the low half of `st_dev`.
    pub device_minor: u32,
    /// Mount identifier in this mount namespace, or zero when unavailable.
    pub mount_id: u64,
    /// Inode number within the mounted filesystem.
    pub inode: u64,
}

impl FileIdentity {
    /// Describe a file from metadata already in hand, or nothing when the
    /// filesystem reports no modification time to compare against later.
    pub fn describe(path: &Path, metadata: &std::fs::Metadata) -> Option<Self> {
        Some(Self {
            path: path.to_path_buf(),
            len: metadata.len(),
            modified: metadata.modified().ok()?,
            changed: change_token(metadata),
            object: metadata_object_identity(metadata),
        })
    }

    /// Describe a file only when metadata can safely stand in for its digest.
    ///
    /// Linux NFS may revise cached change times without a write, so it uses a
    /// cached object/length/mtime identity that deliberately omits ctime. This
    /// is the same timestamp-freshness contract as Cargo and [`VerifiedBlob`];
    /// forcing a server round trip per input would serialize large NFS builds.
    /// If the kernel cannot supply every required field, callers hash instead.
    pub fn for_digest_cache(path: &Path, metadata: &std::fs::Metadata) -> io::Result<Option<Self>> {
        digest_cache_identity(
            path,
            metadata,
            metadata_identity_is_unreliable(path, metadata)?,
        )
    }

    /// Whether the file at this identity's path still has exactly this
    /// identity, so the digest recorded against it still describes the bytes on
    /// disk without reading them again.
    ///
    /// Length alone would miss an overwrite that keeps the size, and the
    /// modification time can be put back by whoever rewrote the file. The
    /// change time cannot be set from user space, so where the platform reports
    /// one a rewrite that restores the old modification time still shows. A
    /// file that has vanished is an error rather than a change, so the caller
    /// can tell the two apart.
    pub fn still_describes(&self) -> std::io::Result<bool> {
        let metadata = std::fs::metadata(&self.path)?;
        Ok(Self::for_digest_cache(&self.path, &metadata)?.as_ref() == Some(self))
    }

    /// Whether this identity is strong enough to stand in for a second read.
    pub fn can_skip_content_verification(&self) -> bool {
        self.changed.is_some() || self.object.is_some()
    }
}

/// A pre-operation snapshot that can prove whether a file's contents changed.
///
/// Most filesystems provide a stable metadata-change token, so the inexpensive
/// identity is sufficient once its timestamp tick has settled. Recently written
/// files carry a content digest alongside their metadata to detect another
/// write within that tick.
/// Linux NFS can reconcile client and server timestamps after a writer has
/// closed the file, making two metadata reads disagree without any intervening
/// write. Those files carry a content digest instead,
/// while retaining length and file identity to detect replacement. Callers use
/// the same comparison either way and do not need to know which filesystem
/// supplied the file.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct FileSnapshot {
    identity: FileIdentity,
    content: Option<CacheDigest>,
    require_full_identity: bool,
}

impl FileSnapshot {
    /// Capture the strongest comparison the file's filesystem can support.
    pub fn capture(path: &Path) -> io::Result<Option<Self>> {
        Self::capture_with_cache(path, &NoFileDigestCache)
    }

    /// Capture a snapshot while reusing or publishing a content digest through
    /// the session ledger when the filesystem or a recent write requires one.
    pub fn capture_with_cache(
        path: &Path,
        digests: &dyn FileDigestCache,
    ) -> io::Result<Option<Self>> {
        let observed_at = SystemTime::now();
        let metadata = std::fs::metadata(path)?;
        let unreliable = metadata_identity_is_unreliable(path, &metadata)?;
        capture_file_snapshot(path, digests, unreliable, unreliable, metadata, observed_at)
    }

    /// Capture a content-backed snapshot whatever the filesystem, for a file
    /// whose metadata is expected to change without its bytes.
    ///
    /// A compiler artifact is one: rustc hardlinks the `.rmeta` it just
    /// emitted into its incremental session directory when it finishes, which
    /// bumps the change token of a file a pipelined dependent is already
    /// reading. Comparing the bytes, length, and file object instead accepts
    /// that and still rejects a replacement or a rewrite.
    pub fn capture_content_with_cache(
        path: &Path,
        digests: &dyn FileDigestCache,
    ) -> io::Result<Option<Self>> {
        let observed_at = SystemTime::now();
        let metadata = std::fs::metadata(path)?;
        // The ledger identity still follows the filesystem: recording under
        // the ordinary identity and looking up under the NFS one would miss
        // every time, and every dependent would re-hash the artifact.
        let unreliable = metadata_identity_is_unreliable(path, &metadata)?;
        capture_file_snapshot(path, digests, unreliable, true, metadata, observed_at)
    }

    /// Whether `identity` and `content` still describe this snapshot.
    ///
    /// Recently written local sources require both their full metadata
    /// identity and content digest to match. NFS and explicitly requested
    /// content snapshots ignore both timestamps: they can change without a
    /// content change. For those timestamp-tolerant snapshots, matching digests
    /// cannot detect an intervening edit that restores the original bytes.
    /// Digest-cache lookup still uses timestamps to decide when a digest needs
    /// refreshing; this comparison performs no additional reads.
    pub fn matches(&self, identity: Option<&FileIdentity>, content: &CacheDigest) -> bool {
        self.content.as_ref().map_or_else(
            || identity == Some(&self.identity),
            |before| {
                before == content
                    && identity.is_some_and(|after| {
                        if self.require_full_identity {
                            &self.identity == after
                        } else {
                            self.identity.path == after.path
                                && self.identity.len == after.len
                                && self.identity.object == after.object
                        }
                    })
            },
        )
    }

    /// Whether a mismatch proves the file's contents changed.
    pub fn proves_content_change(&self) -> bool {
        self.content.is_some() || self.identity.changed.is_some()
    }
}

impl From<FileIdentity> for FileSnapshot {
    fn from(identity: FileIdentity) -> Self {
        Self {
            identity,
            content: None,
            require_full_identity: true,
        }
    }
}

#[cfg(target_os = "linux")]
fn metadata_identity_is_unreliable(path: &Path, metadata: &std::fs::Metadata) -> io::Result<bool> {
    use std::mem::MaybeUninit;
    use std::os::unix::ffi::OsStrExt as _;
    use std::os::unix::fs::MetadataExt as _;

    static FILESYSTEMS: OnceLock<Mutex<std::collections::BTreeMap<u64, bool>>> = OnceLock::new();
    let filesystems = FILESYSTEMS.get_or_init(|| Mutex::new(std::collections::BTreeMap::new()));
    if let Some(unreliable) = filesystems.lock().unwrap().get(&metadata.dev()).copied() {
        return Ok(unreliable);
    }

    let path = std::ffi::CString::new(path.as_os_str().as_bytes())
        .map_err(|_| io::Error::new(io::ErrorKind::InvalidInput, "path contains a NUL byte"))?;
    let mut status = MaybeUninit::<libc::statfs>::zeroed();
    // SAFETY: `path` is NUL-terminated and `status` points to writable,
    // correctly sized storage. statfs initializes it before returning success.
    let result = unsafe { libc::statfs(path.as_ptr(), status.as_mut_ptr()) };
    if result != 0 {
        return Err(io::Error::last_os_error());
    }
    // SAFETY: statfs returned success and initialized `status`.
    let status = unsafe { status.assume_init() };
    // libc gives NFS_SUPER_MAGIC a different signedness from statfs::f_type on musl.
    let unreliable = status.f_type == 0x6969;
    filesystems
        .lock()
        .unwrap()
        .insert(metadata.dev(), unreliable);
    Ok(unreliable)
}

#[cfg(not(target_os = "linux"))]
fn metadata_identity_is_unreliable(
    _path: &Path,
    _metadata: &std::fs::Metadata,
) -> io::Result<bool> {
    Ok(false)
}

/// `unreliable` selects the ledger identity the filesystem calls for, as
/// [`digest_file_validated`] does when it records one; `content_identity`
/// asks for bytes to be compared while tolerating timestamp changes.
fn capture_file_snapshot(
    path: &Path,
    digests: &dyn FileDigestCache,
    unreliable: bool,
    content_identity: bool,
    metadata: std::fs::Metadata,
    observed_at: SystemTime,
) -> io::Result<Option<FileSnapshot>> {
    let cache_identity = digest_cache_identity(path, &metadata, unreliable)?;
    let Some(identity) = cache_identity
        .clone()
        .or_else(|| FileIdentity::describe(path, &metadata))
    else {
        return Ok(None);
    };
    // Equal metadata cannot vouch for bytes written within the same timestamp
    // tick. Such snapshots need content, and a ledger hit under that identity
    // could describe bytes from an earlier write in the tick.
    let racy = identity_is_racy(&identity, observed_at);
    let cache_identity = cache_identity.filter(|_| !racy);
    let (identity, content) = if content_identity || racy {
        let resolved = cache_identity
            .as_ref()
            .and_then(|identity| {
                digests
                    .resolve(FileDigestScope::Content, std::slice::from_ref(identity))
                    .pop()
            })
            .unwrap_or(FileDigestResolution::Unresolved);
        let cached = match resolved {
            FileDigestResolution::Digest(digest)
                if cache_identity.as_ref().is_some_and(|identity| {
                    identity.len == digest.size && identity.still_describes().unwrap_or(false)
                }) =>
            {
                Some(digest)
            }
            FileDigestResolution::Digest(_)
            | FileDigestResolution::EmbeddedTimestampMacro
            | FileDigestResolution::Unresolved => None,
        };
        if let Some(digest) = cached {
            (identity, Some(digest))
        } else {
            let observed = digest_file_validated(FileDigestScope::Content, path)?
                .ok_or_else(|| io::Error::other("could not establish a file digest identity"))?;
            let digest = observed
                .resolution
                .into_digest()
                .ok_or_else(|| io::Error::other("content digest resolution returned no digest"))?;
            if let Some(file) = observed.cache_identity
                && file.len == digest.size
            {
                digests.record(
                    FileDigestScope::Content,
                    vec![RecordedFileDigest {
                        file,
                        digest: digest.clone(),
                    }],
                );
            }
            (observed.identity, Some(digest))
        }
    } else {
        (identity, None)
    };
    Ok(Some(FileSnapshot {
        identity,
        content,
        require_full_identity: !unreliable && !content_identity,
    }))
}

fn digest_cache_identity(
    path: &Path,
    metadata: &std::fs::Metadata,
    unreliable: bool,
) -> io::Result<Option<FileIdentity>> {
    if unreliable {
        nfs_file_identity(path)
    } else {
        Ok(FileIdentity::describe(path, metadata))
    }
}

#[cfg(target_os = "linux")]
fn nfs_file_identity(path: &Path) -> io::Result<Option<FileIdentity>> {
    use std::mem::MaybeUninit;
    use std::os::unix::ffi::OsStrExt as _;

    let path = std::ffi::CString::new(path.as_os_str().as_bytes())
        .map_err(|_| io::Error::new(io::ErrorKind::InvalidInput, "path contains a NUL byte"))?;
    let mut status = MaybeUninit::<LinuxStatx>::zeroed();
    // SAFETY: `path` is NUL-terminated and `status` names writable storage.
    let result = unsafe {
        libc::syscall(
            libc::SYS_statx,
            libc::AT_FDCWD,
            path.as_ptr(),
            // AT_STATX_DONT_SYNC: the surrounding build already trusts cached
            // mtimes, and FORCE_SYNC costs one NFS RPC for every dependency
            // edge rather than every distinct file.
            0x4000,
            STATX_IDENTITY_MASK,
            status.as_mut_ptr(),
        )
    };
    if result != 0 {
        let error = io::Error::last_os_error();
        return match error.raw_os_error() {
            Some(libc::ENOSYS | libc::EINVAL | libc::EOPNOTSUPP) => Ok(None),
            _ => Err(error),
        };
    }
    // SAFETY: statx returned success and initialized `status`.
    let status = unsafe { status.assume_init() };
    nfs_identity_from_statx(
        Path::new(std::ffi::OsStr::from_bytes(path.to_bytes())),
        &status,
    )
}

#[cfg(target_os = "linux")]
fn nfs_identity_from_statx(path: &Path, status: &LinuxStatx) -> io::Result<Option<FileIdentity>> {
    if status.mask & STATX_IDENTITY_MASK != STATX_IDENTITY_MASK
        || status.modified.nanos >= 1_000_000_000
    {
        return Ok(None);
    }
    let modified = system_time(status.modified.seconds, status.modified.nanos)?;
    Ok(Some(FileIdentity {
        path: path.to_path_buf(),
        len: status.size,
        modified,
        changed: None,
        object: Some(FileObjectIdentity {
            device_major: status.device_major,
            device_minor: status.device_minor,
            mount_id: status.mount_id,
            inode: status.inode,
        }),
    }))
}

#[cfg(not(target_os = "linux"))]
fn nfs_file_identity(_path: &Path) -> io::Result<Option<FileIdentity>> {
    Ok(None)
}

#[cfg(target_os = "linux")]
fn system_time(seconds: i64, nanos: u32) -> io::Result<SystemTime> {
    if seconds >= 0 {
        SystemTime::UNIX_EPOCH.checked_add(std::time::Duration::new(seconds as u64, nanos))
    } else {
        SystemTime::UNIX_EPOCH
            .checked_sub(std::time::Duration::from_secs(seconds.unsigned_abs()))
            .and_then(|time| time.checked_add(std::time::Duration::from_nanos(nanos.into())))
    }
    .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "file timestamp is out of range"))
}

/// The metadata-change time as an opaque token, where the platform has one.
#[cfg(unix)]
fn change_token(metadata: &std::fs::Metadata) -> Option<(i64, i64)> {
    use std::os::unix::fs::MetadataExt;
    Some((metadata.ctime(), metadata.ctime_nsec()))
}

/// Windows reports creation rather than metadata-change time, which a rewrite
/// does not move, so no token is better than a misleading one.
#[cfg(not(unix))]
fn change_token(_metadata: &std::fs::Metadata) -> Option<(i64, i64)> {
    None
}

#[cfg(unix)]
fn metadata_object_identity(metadata: &std::fs::Metadata) -> Option<FileObjectIdentity> {
    use std::os::unix::fs::MetadataExt as _;

    let device = metadata.dev();
    Some(FileObjectIdentity {
        device_major: (device >> 32) as u32,
        device_minor: device as u32,
        mount_id: 0,
        inode: metadata.ino(),
    })
}

#[cfg(not(unix))]
fn metadata_object_identity(_metadata: &std::fs::Metadata) -> Option<FileObjectIdentity> {
    None
}

/// A file digest recorded against the identity it was read under.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct RecordedFileDigest {
    /// Identity of the file when its contents were hashed.
    pub file: FileIdentity,
    /// Digest of those contents.
    pub digest: CacheDigest,
}

/// What a recorded file digest may stand in for.
///
/// Adapters prove different things when they read a file: the cc adapter's
/// input scan also establishes that a source embeds no timestamp macro, which
/// a digest recorded by the rustc adapter never checked. Scoping the ledger
/// keeps one adapter's shortcut from resting on a property another adapter
/// never established.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum FileDigestScope {
    /// The digest describes the file's contents and nothing more.
    Content,
    /// The digest describes a cc compiler input that also passed the
    /// timestamp-macro scan.
    CcInput,
}

/// The outcome of resolving one file under a digest scope.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum FileDigestResolution {
    /// The file's content digest, with any scope-specific checks complete.
    Digest(CacheDigest),
    /// A C/C++ input contains a time-dependent preprocessor macro.
    EmbeddedTimestampMacro,
    /// No shared resolver was available; the caller must read the file.
    Unresolved,
}

/// A file digest tied to the file object and pathname observed while reading.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ValidatedFileDigest {
    /// Scope-specific outcome of reading the file.
    pub resolution: FileDigestResolution,
    /// Identity of the pathname after the validated read.
    pub identity: FileIdentity,
    /// Identity suitable for reuse in the digest ledger, when available.
    pub cache_identity: Option<FileIdentity>,
    /// The file was written too recently for its digest to be reused, so
    /// `cache_identity` was withheld. The digest is right for this read only.
    pub racy: bool,
}

impl FileDigestResolution {
    /// Extract the digest when resolution succeeded.
    pub fn into_digest(self) -> Option<CacheDigest> {
        match self {
            Self::Digest(digest) => Some(digest),
            Self::EmbeddedTimestampMacro | Self::Unresolved => None,
        }
    }
}

/// Read and hash a file once, applying the checks required by `scope` in that
/// same pass.
pub fn digest_file(scope: FileDigestScope, path: &Path) -> io::Result<FileDigestResolution> {
    let file = std::fs::File::open(path)?;
    digest_reader(scope, &file).map(|(resolution, _)| resolution)
}

/// Read and hash a regular file only when the open handle remains stable and
/// the pathname still names that same file after the read.
pub fn digest_file_validated(
    scope: FileDigestScope,
    path: &Path,
) -> io::Result<Option<ValidatedFileDigest>> {
    // Taken before the read: a slow read must not make a file look settled
    // that was still inside its timestamp tick when the first byte was read.
    let started = SystemTime::now();
    let file = std::fs::File::open(path)?;
    let before = file.metadata()?;
    if !before.is_file() {
        return Err(io::Error::new(
            io::ErrorKind::InvalidInput,
            format!("digest path is not a regular file: {}", path.display()),
        ));
    }
    let (resolution, size) = digest_reader(scope, &file)?;
    let after = file.metadata()?;
    let current_metadata = std::fs::metadata(path)?;
    let unreliable = metadata_identity_is_unreliable(path, &current_metadata)?;
    let cache_identity = digest_cache_identity(path, &current_metadata, unreliable)?;
    let identity = cache_identity
        .clone()
        .or_else(|| FileIdentity::describe(path, &current_metadata));
    let Some(identity) = identity else {
        return Ok(None);
    };
    if !metadata_names_same_file(&after, &current_metadata) {
        return Err(io::Error::new(
            io::ErrorKind::WouldBlock,
            "file was replaced while its contents were being read",
        ));
    }
    if !handle_and_path_are_stable(path, &before, &after, &identity, unreliable, size)? {
        return Err(io::Error::new(
            io::ErrorKind::WouldBlock,
            "file changed while its contents were being read",
        ));
    }
    // A digest recorded this close to the file's last write cannot be trusted
    // later: a second write in the same timestamp tick leaves the identity
    // unchanged, so the record would keep answering for bytes it never read.
    // The caller still gets the digest; it just is not offered for reuse.
    let racy = cache_identity
        .as_ref()
        .is_some_and(|identity| identity_is_racy(identity, started));
    let cache_identity = cache_identity.filter(|_| !racy);
    Ok(Some(ValidatedFileDigest {
        resolution,
        identity,
        cache_identity,
        racy,
    }))
}

/// Timestamp tick of a filesystem that keeps sub-second times. The kernel
/// stamps files from a coarse clock that advances once per scheduler tick,
/// at most 10 ms, so twice that leaves a margin.
const FINE_TIMESTAMP_RACE_WINDOW: std::time::Duration = std::time::Duration::from_millis(20);

/// Timestamp tick of a filesystem that keeps whole seconds (ext3, HFS+) or two
/// (FAT).
const COARSE_TIMESTAMP_RACE_WINDOW: std::time::Duration = std::time::Duration::from_secs(2);

/// Whether the file was last written so recently that another write could
/// still land in the timestamp tick the identity records.
///
/// The same "racily clean" hazard git guards its index against: metadata that
/// is equal before and after cannot show a write that left it unchanged, so a
/// digest is only reusable once the file's last change is older than the
/// filesystem's timestamp resolution when the read began (`observed_at`). A timestamp in
/// the future is treated as racy too.
fn identity_is_racy(identity: &FileIdentity, observed_at: SystemTime) -> bool {
    let nanos = |time: SystemTime| {
        time.duration_since(SystemTime::UNIX_EPOCH)
            .map_or(0, |since| since.subsec_nanos())
    };
    let coarse = nanos(identity.modified) == 0
        && identity
            .changed
            .is_none_or(|(_, changed_nanos)| changed_nanos == 0);
    let window = if coarse {
        COARSE_TIMESTAMP_RACE_WINDOW
    } else {
        FINE_TIMESTAMP_RACE_WINDOW
    };
    let recent = |time: SystemTime| {
        observed_at
            .duration_since(time)
            .map_or(true, |age| age < window)
    };
    recent(identity.modified)
        || identity.changed.is_some_and(|(seconds, changed_nanos)| {
            u64::try_from(seconds).map_or(true, |seconds| {
                recent(
                    SystemTime::UNIX_EPOCH
                        + std::time::Duration::new(seconds, changed_nanos.try_into().unwrap_or(0)),
                )
            })
        })
}

fn digest_reader(
    scope: FileDigestScope,
    file: &std::fs::File,
) -> io::Result<(FileDigestResolution, u64)> {
    let mut reader = std::io::BufReader::new(file);
    let mut hasher = blake3::Hasher::new();
    let mut size = 0_u64;
    let longest_macro = TIMESTAMP_MACROS
        .iter()
        .map(|macro_name| macro_name.len())
        .max()
        .unwrap_or_default();
    let mut window = Vec::with_capacity(DIGEST_BUFFER_BYTES + longest_macro);
    let mut chunk = vec![0_u8; DIGEST_BUFFER_BYTES];
    let mut found_timestamp_macro = false;
    loop {
        let read = reader.read(&mut chunk)?;
        if read == 0 {
            break;
        }
        hasher.update(&chunk[..read]);
        size = size
            .checked_add(read as u64)
            .ok_or_else(|| io::Error::other("file length overflowed u64"))?;
        if scope == FileDigestScope::CcInput && !found_timestamp_macro {
            window.extend_from_slice(&chunk[..read]);
            found_timestamp_macro = TIMESTAMP_MACROS
                .iter()
                .any(|macro_name| memchr::memmem::find(&window, macro_name).is_some());
            let keep = window.len().saturating_sub(longest_macro.saturating_sub(1));
            window.drain(..keep);
        }
    }
    if found_timestamp_macro {
        Ok((FileDigestResolution::EmbeddedTimestampMacro, size))
    } else {
        Ok((
            FileDigestResolution::Digest(CacheDigest {
                algorithm: "blake3".into(),
                hash: hasher.finalize().to_hex().to_string(),
                size,
            }),
            size,
        ))
    }
}

fn handle_and_path_are_stable(
    path: &Path,
    before: &std::fs::Metadata,
    after: &std::fs::Metadata,
    current: &FileIdentity,
    unreliable: bool,
    digest_size: u64,
) -> io::Result<bool> {
    if before.len() != after.len() || after.len() != digest_size || current.len != digest_size {
        return Ok(false);
    }
    if !metadata_names_same_file(before, after) {
        return Ok(false);
    }
    if unreliable {
        #[cfg(target_os = "linux")]
        {
            let before_modified = before.modified()?;
            let after_modified = after.modified()?;
            let object_matches = current.object.as_ref().is_none_or(|object| {
                metadata_matches_object(before, object) && metadata_matches_object(after, object)
            });
            return Ok(before_modified == after_modified
                && after_modified == current.modified
                && object_matches);
        }
        #[cfg(not(target_os = "linux"))]
        {
            let _ = (path, before, after, current);
            return Ok(false);
        }
    }
    let before = FileIdentity::describe(path, before);
    let after = FileIdentity::describe(path, after);
    Ok(before.is_some() && before == after && after.as_ref() == Some(current))
}

#[cfg(unix)]
fn metadata_names_same_file(left: &std::fs::Metadata, right: &std::fs::Metadata) -> bool {
    use std::os::unix::fs::MetadataExt as _;

    left.dev() == right.dev() && left.ino() == right.ino()
}

#[cfg(not(unix))]
fn metadata_names_same_file(_left: &std::fs::Metadata, _right: &std::fs::Metadata) -> bool {
    true
}

#[cfg(target_os = "linux")]
fn metadata_matches_object(metadata: &std::fs::Metadata, object: &FileObjectIdentity) -> bool {
    use std::os::unix::fs::MetadataExt as _;

    let device = metadata.dev();
    let major = ((device >> 8) & 0x0fff) | ((device >> 32) & 0xfffff000);
    let minor = (device & 0x00ff) | ((device >> 12) & 0xffffff00);
    u32::try_from(major).ok() == Some(object.device_major)
        && u32::try_from(minor).ok() == Some(object.device_minor)
        && metadata.ino() == object.inode
}

/// Recorded digests a session may consult instead of rehashing a file.
///
/// The agent's file-digest ledger answers through this everywhere a caller
/// holds one; the no-op implementation stands in where reuse must not happen,
/// such as under verification.
pub trait FileDigestCache: Send + Sync {
    /// Resolve these identities, coalescing concurrent misses where supported.
    fn resolve(&self, scope: FileDigestScope, files: &[FileIdentity]) -> Vec<FileDigestResolution> {
        self.find(scope, files)
            .into_iter()
            .map(|digest| {
                digest.map_or(
                    FileDigestResolution::Unresolved,
                    FileDigestResolution::Digest,
                )
            })
            .collect()
    }
    /// Recorded digests for these identities, in request order.
    fn find(&self, scope: FileDigestScope, files: &[FileIdentity]) -> Vec<Option<CacheDigest>>;
    /// Record digests for files read under these identities.
    fn record(&self, scope: FileDigestScope, entries: Vec<RecordedFileDigest>);
}

/// A [`FileDigestCache`] that remembers nothing and finds nothing.
pub struct NoFileDigestCache;

impl FileDigestCache for NoFileDigestCache {
    fn resolve(
        &self,
        _scope: FileDigestScope,
        files: &[FileIdentity],
    ) -> Vec<FileDigestResolution> {
        vec![FileDigestResolution::Unresolved; files.len()]
    }

    fn find(&self, _scope: FileDigestScope, files: &[FileIdentity]) -> Vec<Option<CacheDigest>> {
        vec![None; files.len()]
    }

    fn record(&self, _scope: FileDigestScope, _entries: Vec<RecordedFileDigest>) {}
}

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

    #[cfg(target_os = "linux")]
    #[test]
    fn an_nfs_snapshot_consults_the_session_digest_ledger() {
        struct CountingCache(std::sync::atomic::AtomicUsize);
        impl FileDigestCache for CountingCache {
            fn find(
                &self,
                scope: FileDigestScope,
                files: &[FileIdentity],
            ) -> Vec<Option<CacheDigest>> {
                assert_eq!(scope, FileDigestScope::Content);
                assert_eq!(files.len(), 1);
                assert_eq!(files[0].changed, None);
                assert!(files[0].object.is_some());
                self.0.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                vec![None]
            }
            fn record(&self, _scope: FileDigestScope, _entries: Vec<RecordedFileDigest>) {}
        }
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rlib");
        std::fs::write(&path, b"nfs bytes").unwrap();
        if nfs_file_identity(&path).unwrap().is_none() {
            eprintln!("skipping: statx cannot supply the NFS ledger identity");
            return;
        }
        let cache = CountingCache(std::sync::atomic::AtomicUsize::new(0));
        let metadata = std::fs::metadata(&path).unwrap();
        // Settled: a racy identity skips the ledger by design.
        let observed_at = metadata.modified().unwrap() + std::time::Duration::from_secs(3);
        let snapshot = capture_file_snapshot(&path, &cache, true, true, metadata, observed_at)
            .unwrap()
            .unwrap();
        assert_eq!(snapshot.content, Some(CacheDigest::blake3(b"nfs bytes")));
        assert_eq!(cache.0.load(std::sync::atomic::Ordering::Relaxed), 1);
    }

    #[test]
    fn timestamp_macros_are_detected_at_every_chunk_boundary_split() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.h");
        for macro_name in TIMESTAMP_MACROS {
            // Include a macro starting at the boundary and one ending there.
            for split in 0..=macro_name.len() {
                let mut bytes = vec![b'x'; DIGEST_BUFFER_BYTES - split];
                bytes.extend_from_slice(macro_name);
                std::fs::write(&path, &bytes).unwrap();
                assert_eq!(
                    digest_file(FileDigestScope::CcInput, &path).unwrap(),
                    FileDigestResolution::EmbeddedTimestampMacro,
                    "macro {macro_name:?}, split {split}"
                );
                // The same bytes remain hashable outside the C/C++ scope.
                assert_eq!(
                    digest_file(FileDigestScope::Content, &path).unwrap(),
                    FileDigestResolution::Digest(CacheDigest::blake3(&bytes))
                );
            }
        }
    }

    #[test]
    fn timestamp_scan_accepts_empty_short_and_near_match_inputs() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.h");
        for bytes in [
            b"".as_slice(),
            b"_",
            b"__DATE_ __TIME_ __TIMESTAMP_",
            b"__date__ __time__ __timestamp__",
            &[0xff, 0, b'_', b'_'],
        ] {
            std::fs::write(&path, bytes).unwrap();
            assert_eq!(
                digest_file(FileDigestScope::CcInput, &path).unwrap(),
                FileDigestResolution::Digest(CacheDigest::blake3(bytes))
            );
        }
    }

    /// rustc links a finished `.rmeta` into its incremental directory, which
    /// changes the file's metadata but not its bytes. A metadata snapshot
    /// cannot tell that from a rewrite; a content snapshot can, and still
    /// notices when the bytes do change.
    #[cfg(unix)]
    #[test]
    fn a_content_snapshot_survives_a_hardlink_and_a_metadata_one_does_not() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("libdep.rmeta");
        std::fs::write(&path, b"metadata").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap() + std::time::Duration::from_secs(3);
        let by_metadata = capture_file_snapshot(
            &path,
            &NoFileDigestCache,
            false,
            false,
            metadata,
            observed_at,
        )
        .unwrap()
        .unwrap();
        let by_content = FileSnapshot::capture_content_with_cache(&path, &NoFileDigestCache)
            .unwrap()
            .unwrap();
        let digest = CacheDigest::blake3_file(&path).unwrap();
        // A new ctime, so the change token cannot be mistaken for the old one.
        std::thread::sleep(std::time::Duration::from_millis(20));
        std::fs::hard_link(&path, directory.path().join("metadata.rmeta")).unwrap();
        let after = FileIdentity::describe(&path, &std::fs::metadata(&path).unwrap()).unwrap();

        assert!(by_content.proves_content_change());
        assert!(by_content.matches(Some(&after), &digest));
        if by_metadata.proves_content_change() {
            assert!(!by_metadata.matches(Some(&after), &digest));
        }

        std::fs::write(&path, b"rewritten").unwrap();
        let rewritten = CacheDigest::blake3_file(&path).unwrap();
        let after = FileIdentity::describe(&path, &std::fs::metadata(&path).unwrap()).unwrap();
        assert!(!by_content.matches(Some(&after), &rewritten));
    }

    #[cfg(unix)]
    struct RecordingDigestCache {
        answer: Option<CacheDigest>,
        recorded: std::sync::Mutex<Vec<RecordedFileDigest>>,
    }

    #[cfg(unix)]
    impl FileDigestCache for RecordingDigestCache {
        fn find(
            &self,
            _scope: FileDigestScope,
            files: &[FileIdentity],
        ) -> Vec<Option<CacheDigest>> {
            assert_eq!(files.len(), 1);
            vec![self.answer.clone()]
        }

        fn record(&self, _scope: FileDigestScope, entries: Vec<RecordedFileDigest>) {
            self.recorded.lock().unwrap().extend(entries);
        }
    }

    #[cfg(unix)]
    #[test]
    fn content_snapshot_records_a_replacement_under_its_own_identity() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rlib");
        let replacement = directory.path().join("replacement.rlib");
        std::fs::write(&path, b"original....").unwrap();
        std::fs::File::options()
            .write(true)
            .open(&path)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(SystemTime::UNIX_EPOCH))
            .unwrap();
        std::fs::write(&replacement, b"replacement!").unwrap();
        std::fs::File::options()
            .write(true)
            .open(&replacement)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(SystemTime::UNIX_EPOCH))
            .unwrap();
        let before = std::fs::metadata(&path).unwrap();
        let old_identity = FileIdentity::describe(&path, &before).unwrap();
        std::fs::remove_file(&path).unwrap();
        std::fs::rename(&replacement, &path).unwrap();
        let cache = RecordingDigestCache {
            answer: None,
            recorded: std::sync::Mutex::new(Vec::new()),
        };

        settle_past_timestamp_race();
        let snapshot = capture_file_snapshot(&path, &cache, true, true, before, SystemTime::now())
            .unwrap()
            .unwrap();
        let current = FileIdentity::for_digest_cache(&path, &std::fs::metadata(&path).unwrap())
            .unwrap()
            .unwrap();
        let digest = CacheDigest::blake3_file(&path).unwrap();

        assert_eq!(current.modified, old_identity.modified);
        assert_ne!(current.object, old_identity.object);
        assert_ne!(current, old_identity);
        assert!(snapshot.matches(Some(&current), &digest));
        let recorded = cache.recorded.lock().unwrap();
        assert_eq!(recorded.len(), 1);
        assert_eq!(recorded[0].file, current);
        assert_eq!(recorded[0].digest, digest);
    }

    #[cfg(unix)]
    #[test]
    fn content_snapshot_rejects_a_stale_digest_cache_hit() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rlib");
        let replacement = directory.path().join("replacement.rlib");
        std::fs::write(&path, b"original....").unwrap();
        std::fs::File::options()
            .write(true)
            .open(&path)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(SystemTime::UNIX_EPOCH))
            .unwrap();
        let before = std::fs::metadata(&path).unwrap();
        std::fs::write(&replacement, b"replacement!").unwrap();
        std::fs::File::options()
            .write(true)
            .open(&replacement)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(SystemTime::UNIX_EPOCH))
            .unwrap();
        std::fs::remove_file(&path).unwrap();
        std::fs::rename(&replacement, &path).unwrap();
        let cache = RecordingDigestCache {
            answer: Some(CacheDigest::blake3(b"original....")),
            recorded: std::sync::Mutex::new(Vec::new()),
        };

        settle_past_timestamp_race();
        let snapshot = capture_file_snapshot(&path, &cache, true, true, before, SystemTime::now())
            .unwrap()
            .unwrap();
        let current = FileIdentity::for_digest_cache(&path, &std::fs::metadata(&path).unwrap())
            .unwrap()
            .unwrap();
        let digest = CacheDigest::blake3_file(&path).unwrap();

        assert!(snapshot.matches(Some(&current), &digest));
        assert_ne!(snapshot.content, Some(CacheDigest::blake3(b"original....")));
        assert_eq!(cache.recorded.lock().unwrap()[0].digest, digest);
    }

    #[cfg(unix)]
    #[test]
    fn a_replaced_path_does_not_validate_the_open_handle() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rlib");
        let replacement = directory.path().join("replacement.rlib");
        std::fs::write(&path, b"old bytes").unwrap();
        std::fs::write(&replacement, b"new bytes").unwrap();

        let file = std::fs::File::open(&path).unwrap();
        let before = file.metadata().unwrap();
        std::fs::rename(&replacement, &path).unwrap();
        let (_, size) = digest_reader(FileDigestScope::Content, &file).unwrap();
        let after = file.metadata().unwrap();
        let current_metadata = std::fs::metadata(&path).unwrap();
        let current = FileIdentity::describe(&path, &current_metadata).unwrap();

        assert!(!metadata_names_same_file(&after, &current_metadata));
        assert!(
            !handle_and_path_are_stable(&path, &before, &after, &current, false, size).unwrap()
        );
    }

    #[test]
    fn an_identity_describes_the_file_until_it_is_written_or_removed() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let identity = FileIdentity::describe(&path, &std::fs::metadata(&path).unwrap()).unwrap();
        assert!(identity.still_describes().unwrap());

        std::thread::sleep(std::time::Duration::from_millis(20));
        std::fs::write(&path, b"fn main() { }").unwrap();
        assert!(!identity.still_describes().unwrap());

        std::fs::remove_file(&path).unwrap();
        assert!(identity.still_describes().is_err());
    }

    #[test]
    fn a_racy_snapshot_rejects_a_same_length_rewrite_with_identical_metadata() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap();
        let snapshot = capture_file_snapshot(
            &path,
            &NoFileDigestCache,
            false,
            false,
            metadata,
            observed_at,
        )
        .unwrap()
        .unwrap();
        let identity = snapshot.identity.clone();
        let original = CacheDigest::blake3_file(&path).unwrap();
        assert!(snapshot.matches(Some(&identity), &original));

        std::fs::write(&path, b"fn main(){ }").unwrap();
        let rewritten = CacheDigest::blake3_file(&path).unwrap();
        assert_eq!(original.size, rewritten.size);
        assert_ne!(original, rewritten);
        // Model a filesystem tick collision independently of the host's
        // actual timestamp resolution: every metadata field is unchanged.
        assert!(!snapshot.matches(Some(&identity), &rewritten));
        assert!(snapshot.proves_content_change());
    }

    #[cfg(unix)]
    #[test]
    fn a_racy_local_snapshot_rejects_an_edit_that_restores_the_original_bytes() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap();
        let snapshot = capture_file_snapshot(
            &path,
            &NoFileDigestCache,
            false,
            false,
            metadata,
            observed_at,
        )
        .unwrap()
        .unwrap();
        let original = CacheDigest::blake3_file(&path).unwrap();
        assert!(snapshot.matches(Some(&snapshot.identity), &original));

        std::fs::write(&path, b"fn main(){ }").unwrap();
        std::fs::write(&path, b"fn main() {}").unwrap();
        let restored = CacheDigest::blake3_file(&path).unwrap();
        assert_eq!(original, restored);
        // Model a changed ctime independently of the host's timestamp tick.
        // The compiler may have consumed the intervening bytes even though
        // discovery sees the original content and mtime again.
        let mut after = snapshot.identity.clone();
        let (seconds, nanos) = after.changed.unwrap();
        after.changed = Some((seconds + 1, nanos));
        assert!(!snapshot.matches(Some(&after), &restored));
    }

    #[test]
    fn a_settled_snapshot_keeps_the_metadata_only_path() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap() + std::time::Duration::from_secs(3);
        // Once metadata has been read, this path must not open the file or
        // consult the content ledger just to capture a settled snapshot.
        std::fs::remove_file(&path).unwrap();
        let snapshot = capture_file_snapshot(
            &path,
            &NoFileDigestCache,
            false,
            false,
            metadata,
            observed_at,
        )
        .unwrap()
        .unwrap();
        assert!(snapshot.content.is_none());
    }

    #[cfg(unix)]
    #[test]
    fn a_racy_snapshot_does_not_reuse_a_digest_with_identical_metadata() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap();
        let cache = RecordingDigestCache {
            answer: Some(CacheDigest::blake3(b"fn main(){ }")),
            recorded: std::sync::Mutex::new(Vec::new()),
        };
        for content_identity in [false, true] {
            let snapshot = capture_file_snapshot(
                &path,
                &cache,
                false,
                content_identity,
                metadata.clone(),
                observed_at,
            )
            .unwrap()
            .unwrap();
            assert_eq!(snapshot.content, Some(CacheDigest::blake3(b"fn main() {}")));
        }
    }

    #[test]
    fn a_metadata_snapshot_detects_a_metadata_change() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();
        let observed_at = metadata.modified().unwrap() + std::time::Duration::from_secs(3);
        let snapshot = capture_file_snapshot(
            &path,
            &NoFileDigestCache,
            false,
            false,
            metadata,
            observed_at,
        )
        .unwrap()
        .unwrap();

        std::fs::File::options()
            .write(true)
            .open(&path)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(SystemTime::UNIX_EPOCH))
            .unwrap();
        let identity = FileIdentity::describe(&path, &std::fs::metadata(&path).unwrap());
        let digest = CacheDigest::blake3_file(&path).unwrap();

        assert!(!snapshot.matches(identity.as_ref(), &digest));
        assert_eq!(snapshot.proves_content_change(), cfg!(unix));
    }

    #[test]
    fn a_content_snapshot_ignores_timestamp_churn_but_detects_changed_bytes() {
        for (unreliable, content_identity) in [(true, false), (true, true), (false, true)] {
            let directory = tempfile::tempdir().unwrap();
            let path = directory.path().join("input.rs");
            std::fs::write(&path, b"fn main() {}").unwrap();
            let snapshot = capture_file_snapshot(
                &path,
                &NoFileDigestCache,
                unreliable,
                content_identity,
                std::fs::metadata(&path).unwrap(),
                SystemTime::now(),
            )
            .unwrap()
            .unwrap();

            let mut identity = snapshot.identity.clone();
            identity.changed = identity
                .changed
                .map(|(seconds, nanos)| (seconds + 1, nanos));
            let digest = CacheDigest::blake3_file(&path).unwrap();
            assert!(snapshot.matches(Some(&identity), &digest));
            assert!(snapshot.proves_content_change());

            identity.modified = SystemTime::UNIX_EPOCH;
            assert!(snapshot.matches(Some(&identity), &digest));

            std::fs::write(&path, b"fn main(){ }").unwrap();
            let identity = FileIdentity::describe(&path, &std::fs::metadata(&path).unwrap());
            let digest = CacheDigest::blake3_file(&path).unwrap();
            assert!(!snapshot.matches(identity.as_ref(), &digest));
        }
    }

    #[test]
    fn an_nfs_object_identity_can_skip_a_second_content_read() {
        let identity = FileIdentity {
            path: PathBuf::from("/nfs/input.rlib"),
            len: 9,
            modified: SystemTime::UNIX_EPOCH,
            changed: None,
            object: Some(FileObjectIdentity {
                device_major: 0,
                device_minor: 42,
                mount_id: 7,
                inode: 99,
            }),
        };
        assert!(identity.can_skip_content_verification());
        let without_object = FileIdentity {
            object: None,
            ..identity
        };
        assert!(!without_object.can_skip_content_verification());
    }

    #[test]
    fn reliable_metadata_keeps_the_native_identity() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("input.rs");
        std::fs::write(&path, b"fn main() {}").unwrap();
        let metadata = std::fs::metadata(&path).unwrap();

        let identity = digest_cache_identity(&path, &metadata, false)
            .unwrap()
            .unwrap();

        assert_eq!(identity.path, path);
        assert_eq!(identity.len, 12);
        assert_eq!(identity.object.is_some(), cfg!(unix));
    }

    #[test]
    fn content_snapshot_ignores_nfs_timestamp_churn_but_not_content_or_identity_changes() {
        let digest = CacheDigest::blake3(b"nfs bytes");
        let identity = FileIdentity {
            path: PathBuf::from("/nfs/input.rlib"),
            len: digest.size,
            modified: SystemTime::UNIX_EPOCH + std::time::Duration::from_secs(10),
            changed: Some((10, 1)),
            object: Some(FileObjectIdentity {
                device_major: 0,
                device_minor: 42,
                mount_id: 7,
                inode: 99,
            }),
        };
        let snapshot = FileSnapshot {
            identity: identity.clone(),
            content: Some(digest.clone()),
            require_full_identity: false,
        };
        // Client/server reconciliation can move either timestamp in either
        // direction. None of these observations proves another write.
        for seconds in [9, 11] {
            let mut after = identity.clone();
            after.changed = Some((seconds, 500));
            after.modified = SystemTime::UNIX_EPOCH + std::time::Duration::new(seconds as u64, 500);
            assert!(snapshot.matches(Some(&after), &digest));
        }

        // Timestamp tolerance does not admit changed bytes, even when their
        // length and all metadata remain the same.
        let changed = CacheDigest::blake3(b"NFS bytes");
        assert_eq!(changed.size, digest.size);
        assert!(!snapshot.matches(Some(&identity), &changed));
        assert!(!snapshot.matches(None, &digest));

        let mut after = identity.clone();
        after.object.as_mut().unwrap().inode += 1;
        assert!(!snapshot.matches(Some(&after), &digest));

        let mut after = identity.clone();
        after.len += 1;
        assert!(!snapshot.matches(Some(&after), &digest));

        let mut after = identity;
        after.path.set_file_name("replacement.rlib");
        assert!(!snapshot.matches(Some(&after), &digest));
    }

    #[cfg(target_os = "linux")]
    #[test]
    fn incomplete_statx_metadata_falls_back_to_content_hashing() {
        // SAFETY: all-zero is a valid unpopulated statx result for this parser.
        let status = unsafe { std::mem::zeroed::<LinuxStatx>() };
        assert!(
            nfs_identity_from_statx(Path::new("/nfs/input.rlib"), &status)
                .unwrap()
                .is_none()
        );
    }

    /// Let a just-written file age past the window in which its digest is
    /// withheld from reuse.
    fn settle_past_timestamp_race() {
        std::thread::sleep(FINE_TIMESTAMP_RACE_WINDOW * 2);
    }

    /// A modification time ahead of the clock is racy however slowly the test
    /// runs, so the withheld case does not depend on how long a stall lasts.
    #[test]
    fn a_digest_of_a_recently_changed_file_is_not_offered_for_reuse() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("val.rs");
        std::fs::write(&path, b"pub const VAL: &str = \"v1\";").unwrap();
        std::fs::File::options()
            .write(true)
            .open(&path)
            .unwrap()
            .set_times(
                std::fs::FileTimes::new()
                    .set_modified(SystemTime::now() + std::time::Duration::from_secs(3600)),
            )
            .unwrap();

        let fresh = digest_file_validated(FileDigestScope::Content, &path)
            .unwrap()
            .unwrap();
        assert!(fresh.resolution.into_digest().is_some());
        assert!(fresh.racy);
        assert_eq!(fresh.cache_identity, None);
    }

    #[test]
    fn a_digest_of_a_settled_file_is_offered_for_reuse() {
        let directory = tempfile::tempdir().unwrap();
        let path = directory.path().join("val.rs");
        std::fs::write(&path, b"pub const VAL: &str = \"v1\";").unwrap();

        settle_past_timestamp_race();
        let settled = digest_file_validated(FileDigestScope::Content, &path)
            .unwrap()
            .unwrap();
        assert!(!settled.racy);
        assert!(settled.cache_identity.is_some());
    }

    #[test]
    fn racy_window_follows_the_timestamp_resolution() {
        let at = |seconds: u64, nanos: u32| {
            SystemTime::UNIX_EPOCH + std::time::Duration::new(seconds, nanos)
        };
        let identity = |modified: SystemTime, changed: Option<(i64, i64)>| FileIdentity {
            path: PathBuf::from("/x"),
            len: 1,
            modified,
            changed,
            object: None,
        };
        let now = at(1_000, 500_000_000);
        let fine = |age_ms: u64| {
            let when = now - std::time::Duration::from_millis(age_ms);
            let since = when.duration_since(SystemTime::UNIX_EPOCH).unwrap();
            identity(
                when,
                Some((since.as_secs() as i64, i64::from(since.subsec_nanos()))),
            )
        };
        assert!(identity_is_racy(&fine(5), now));
        assert!(!identity_is_racy(&fine(25), now));
        // A ctime newer than the mtime counts even when the mtime is old.
        let touched = identity(at(10, 7), Some((1_000, 499_000_000)));
        assert!(identity_is_racy(&touched, now));
        // Whole-second stamps are withheld for two seconds, even without ctime.
        for with_ctime in [false, true] {
            let coarse = |age: u64| {
                identity(
                    at(1_000 - age, 0),
                    with_ctime.then_some((1_000 - age as i64, 0)),
                )
            };
            assert!(identity_is_racy(&coarse(1), at(1_000, 0)));
            assert!(!identity_is_racy(&coarse(3), at(1_000, 0)));
        }
        // A timestamp in the future is racy.
        assert!(identity_is_racy(&identity(at(2_000, 1), None), now));
    }
}