fix(artifacts): harden reconciliation scanner

crates/crank-artifacts/src/housekeeping.rs

@@ -1,12 +1,13 @@
 use std::{
     fmt,
-    os::fd::AsRawFd,
+    os::fd::{AsRawFd, FromRawFd, OwnedFd},
     time::{Duration, SystemTime, UNIX_EPOCH},
 };
 
+use crate::temp_scan::{list_names, valid_temp_name};
 use crate::{
     ArtifactError, ArtifactStore, ReconciliationMutation, ReconciliationNamespace,
-    ReconciliationScan,
+    ReconciliationScan, ReconciliationScanStop,
     store::{
         QUARANTINE_DIR, RootLock, check_file, checkpointed, ensure_root_unchanged, fsync_fd,
         open_existing_dir, open_or_create_dir, rename_no_replace_at, stat_fd, unlinkat,
@@ -16,24 +17,54 @@ use crate::{
 const FINAL_NAMESPACE: u8 = 0;
 const QUARANTINE_NAMESPACE: u8 = 1;
 
-/// Opaque continuation for a bounded reconciliation traversal.  It is valid
+/// Opaque bounded-scan continuation, valid only for an unchanged namespace.
-/// only for the unchanged pinned root from which it was returned.
+/// Discard it after any put, quarantine, delete, or external mutation.
 pub struct ReconciliationCursor {
     root_dev: u64,
     root_ino: u64,
     namespace: u8,
     shard: u8,
-    position: i64,
+    state: ReconciliationCursorState,
 }
 
+enum ReconciliationCursorState {
+    OpenNamespace,
+    ValidateNamespace(OwnedFd),
+    OpenShard(OwnedFd),
+    ValidateShard {
+        namespace: OwnedFd,
+        shard: OwnedFd,
+    },
+    ReadShard {
+        namespace: OwnedFd,
+        shard: OwnedFd,
+        shard_dev: u64,
+        shard_ino: u64,
+        buffer: Vec<u8>,
+        offset: usize,
+        filled: usize,
+        cookie: i64,
+    },
+    SeekShard {
+        namespace: OwnedFd,
+        shard: OwnedFd,
+        shard_dev: u64,
+        shard_ino: u64,
+        buffer: Vec<u8>,
+        cookie: i64,
+    },
+    Transition,
+}
+
+struct ShardIdentity(u64, u64);
+
 impl fmt::Debug for ReconciliationCursor {
     fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
         formatter.write_str("ReconciliationCursor(..)")
     }
 }
 
-/// Opaque, inode-bound evidence for one valid final or quarantined artifact.
+/// Opaque inode-bound evidence that exposes neither digest nor filesystem path.
-/// It intentionally exposes neither its digest nor filesystem path.
 #[derive(Clone)]
 pub struct ReconciliationCandidate {
     root_dev: u64,
@@ -61,8 +92,7 @@ impl ReconciliationCandidate {
     }
 }
 
-/// Opaque, single-use evidence that a particular stale temporary inode was
+/// Opaque single-use evidence for a stale temporary inode under the pinned root.
-/// observed under this store's pinned root. It is not a blob reference.
 #[derive(Debug)]
 pub struct StaleTemp {
     name: String,
@@ -83,8 +113,7 @@ pub struct TempScan {
 }
 
 impl ArtifactStore {
-    /// Streams valid final and quarantine entries with an opaque continuation.
+    /// Streams entries without disclosing paths/digests, with opaque continuation.
-    /// Neither the report nor the returned capabilities disclose paths/digests.
     pub fn scan_reconciliation(
         &self,
         continuation: Option<ReconciliationCursor>,
@@ -92,8 +121,6 @@ impl ArtifactStore {
         result_limit: usize,
     ) -> Result<(ReconciliationScan, Vec<ReconciliationCandidate>), ArtifactError> {
         let root = self.root()?;
-        let _lock = RootLock::shared(root)?;
-        ensure_root_unchanged(root)?;
         let mut cursor = match continuation {
             Some(cursor) => {
                 if cursor.root_dev != root.dev
@@ -104,128 +131,89 @@ impl ArtifactStore {
                 }
                 cursor
             }
-            None => reconciliation_cursor(root.dev, root.ino, FINAL_NAMESPACE, 0, 0),
+            None => reconciliation_cursor(root.dev, root.ino),
         };
-        let mut report = ReconciliationScan::empty(None);
+        let report = ReconciliationScan::empty(None);
-        if scan_budget == 0 || result_limit == 0 {
+        let _lock = match RootLock::shared(root) {
-            report.continuation = Some(cursor);
+            Ok(lock) => lock,
-            return Ok((report, Vec::new()));
+            Err(ArtifactError::Storage) => {
+                return Ok(reconciliation_page(
+                    report,
+                    Vec::new(),
+                    cursor,
+                    ReconciliationScanStop::Retryable,
+                ));
+            }
+            Err(error) => return Err(error),
+        };
+        match ensure_root_unchanged(root) {
+            Ok(()) => {}
+            Err(ArtifactError::Storage) => {
+                return Ok(reconciliation_page(
+                    report,
+                    Vec::new(),
+                    cursor,
+                    ReconciliationScanStop::Retryable,
+                ));
+            }
+            Err(error) => return Err(error),
+        }
+        let mut report = report;
+        if scan_budget == 0 {
+            return Ok(reconciliation_page(
+                report,
+                Vec::new(),
+                cursor,
+                ReconciliationScanStop::ScanBudget,
+            ));
+        }
+        if result_limit == 0 {
+            return Ok(reconciliation_page(
+                report,
+                Vec::new(),
+                cursor,
+                ReconciliationScanStop::ResultLimit,
+            ));
         }
 
-        let mut remaining = scan_budget;
         let mut entries = Vec::new();
         while cursor.namespace <= QUARANTINE_NAMESPACE {
-            if remaining == 0 || entries.len() == result_limit {
+            if entries.len() == result_limit {
-                report.continuation = Some(cursor);
+                return Ok(reconciliation_page(
-                return Ok((report, entries));
+                    report,
+                    entries,
+                    cursor,
+                    ReconciliationScanStop::ResultLimit,
+                ));
             }
-            let namespace_name = namespace_name(cursor.namespace);
+            if report.traversal_syscalls == scan_budget {
-            let namespace = match open_existing_dir(root.fd.as_raw_fd(), namespace_name) {
+                return Ok(reconciliation_page(
-                Ok(fd) => fd,
+                    report,
-                Err(ArtifactError::NotFound) => {
+                    entries,
-                    advance_namespace(&mut cursor);
+                    cursor,
-                    continue;
+                    ReconciliationScanStop::ScanBudget,
-                }
+                ));
-                Err(error) => return Err(error),
+            }
-            };
+            let state = std::mem::replace(&mut cursor.state, ReconciliationCursorState::Transition);
-            loop {
+            match reconciliation_step(root, &mut cursor, state, &mut report, &mut entries) {
-                if remaining == 0 {
+                Ok(()) => {}
-                    report.continuation = Some(cursor);
+                Err((state, ArtifactError::Storage)) => {
-                    return Ok((report, entries));
+                    cursor.state = state;
-                }
+                    return Ok(reconciliation_page(
-                let shard_name = format!("{:02x}", cursor.shard);
+                        report,
-                let shard = match open_existing_dir(namespace.as_raw_fd(), shard_name.as_bytes()) {
+                        entries,
-                    Ok(fd) => fd,
+                        cursor,
-                    Err(ArtifactError::NotFound) => {
+                        ReconciliationScanStop::Retryable,
-                        consume_scan_work(&mut report, &mut remaining);
+                    ));
-                        if cursor.shard == u8::MAX {
+                }
-                            break;
+                Err((_, error)) => return Err(error),
-                        }
-                        advance_shard(&mut cursor);
-                        continue;
-                    }
-                    Err(error) => return Err(error),
-                };
-                let shard_stat = stat_fd(shard.as_raw_fd())?;
-                let mut directory = DirectoryStream::open(shard.as_raw_fd(), cursor.position)?;
-                loop {
-                    if remaining == 0 || entries.len() == result_limit {
-                        report.continuation = Some(cursor);
-                        return Ok((report, entries));
-                    }
-                    let Some((name, after)) = directory.next()? else {
-                        // An empty directory still consumes one bounded
-                        // traversal unit, so a sparse namespace cannot make
-                        // us probe all 256 shards in one page.
-                        consume_scan_work(&mut report, &mut remaining);
-                        break;
-                    };
-                    cursor.position = after;
-                    consume_scan_work(&mut report, &mut remaining);
-                    let Ok(name) = String::from_utf8(name) else {
-                        report.malformed += 1;
-                        continue;
-                    };
-                    if !valid_final_name(&name, &shard_name) {
-                        report.malformed += 1;
-                        continue;
-                    }
-                    #[cfg(debug_assertions)]
-                    let _ = crate::test_support::checkpoint("reconciliation_before_stat");
-                    let stat = match checkpointed("reconciliation_stat", || {
-                        nofollow_stat(shard.as_raw_fd(), &name)
-                    }) {
-                        Ok(stat) => stat,
-                        Err(ArtifactError::NotFound) => {
-                            // The entry was concurrently removed.  It is no
-                            // longer retryable, so the cursor may advance.
-                            continue;
-                        }
-                        Err(ArtifactError::UnsafeRoot) => {
-                            report.unsafe_entries += 1;
-                            continue;
-                        }
-                        Err(error) => return Err(error),
-                    };
-                    if check_file(&stat).is_err() {
-                        report.unsafe_entries += 1;
-                        continue;
-                    }
-                    let namespace = if cursor.namespace == FINAL_NAMESPACE {
-                        report.final_entries += 1;
-                        ReconciliationNamespace::Final
-                    } else {
-                        report.quarantined_entries += 1;
-                        ReconciliationNamespace::Quarantine
-                    };
-                    entries.push(ReconciliationCandidate {
-                        root_dev: root.dev,
-                        root_ino: root.ino,
-                        namespace,
-                        shard: shard_name.clone(),
-                        shard_dev: shard_stat.st_dev,
-                        shard_ino: shard_stat.st_ino,
-                        name,
-                        dev: stat.st_dev,
-                        ino: stat.st_ino,
-                        modified_seconds: stat.st_mtime,
-                        modified_nanoseconds: stat.st_mtime_nsec,
-                    });
-                    // The continuation is deliberately kept at `after`,
-                    // immediately before the next unissued item.
-                }
-                if cursor.shard == u8::MAX {
-                    break;
-                }
-                advance_shard(&mut cursor);
             }
-            advance_namespace(&mut cursor);
         }
+        report.stop = ReconciliationScanStop::Complete;
         Ok((report, entries))
     }
 
-    /// Moves a revalidated final inode into the private quarantine namespace
+    /// Moves a revalidated final inode to quarantine without replacing an inode.
-    /// without replacing any existing quarantine inode.
     pub fn quarantine_reconciliation(
         &self,
         candidate: ReconciliationCandidate,
@@ -271,8 +259,7 @@ impl ArtifactStore {
         }
     }
 
-    /// Idempotently removes a revalidated quarantined inode and fsyncs its
+    /// Idempotently removes a revalidated quarantined inode and fsyncs its directory.
-    /// directory. An fsync ambiguity is explicitly retryable.
     pub fn delete_quarantined_reconciliation(
         &self,
         candidate: ReconciliationCandidate,
@@ -336,8 +323,7 @@ impl ArtifactStore {
             Err(_) => Ok(ReconciliationMutation::Retryable),
         }
     }
-    /// Streams temporary entries, reading at most `scan_budget` directory
+    /// Scans at most `scan_budget` entries and returns `result_limit` stale capabilities.
-    /// entries and returning at most `result_limit` stale capabilities.
     pub fn scan_stale_temps(
         &self,
         grace: Duration,
@@ -409,8 +395,7 @@ impl ArtifactStore {
         Ok((report, result))
     }
 
-    /// Deletes only a previously discovered temporary inode after an exclusive
+    /// Deletes a revalidated stale inode under an exclusive lock, then fsyncs.
-    /// cooperative root lock and inode revalidation, then fsyncs its shard.
     pub fn delete_stale_temp(&self, candidate: StaleTemp) -> Result<(), ArtifactError> {
         let root = self.root()?;
         let _lock = RootLock::exclusive(root)?;
@@ -453,22 +438,433 @@ impl ArtifactStore {
     }
 }
 
-fn reconciliation_cursor(
+fn reconciliation_cursor(root_dev: u64, root_ino: u64) -> ReconciliationCursor {
-    root_dev: u64,
-    root_ino: u64,
-    namespace: u8,
-    shard: u8,
-    position: i64,
-) -> ReconciliationCursor {
     ReconciliationCursor {
         root_dev,
         root_ino,
+        namespace: FINAL_NAMESPACE,
+        shard: 0,
+        state: ReconciliationCursorState::OpenNamespace,
+    }
+}
+
+fn reconciliation_page(
+    mut report: ReconciliationScan,
+    entries: Vec<ReconciliationCandidate>,
+    cursor: ReconciliationCursor,
+    stop: ReconciliationScanStop,
+) -> (ReconciliationScan, Vec<ReconciliationCandidate>) {
+    report.stop = stop;
+    report.continuation = Some(cursor);
+    (report, entries)
+}
+
+fn reconciliation_step(
+    root: &crate::store::Root,
+    cursor: &mut ReconciliationCursor,
+    state: ReconciliationCursorState,
+    report: &mut ReconciliationScan,
+    entries: &mut Vec<ReconciliationCandidate>,
+) -> Result<(), (ReconciliationCursorState, ArtifactError)> {
+    match state {
+        ReconciliationCursorState::OpenNamespace => {
+            count_traversal_syscall(report);
+            match open_traversal_dir(root.fd.as_raw_fd(), namespace_name(cursor.namespace)) {
+                Ok(namespace) => {
+                    cursor.state = ReconciliationCursorState::ValidateNamespace(namespace);
+                }
+                Err(ArtifactError::NotFound) => advance_namespace(cursor),
+                Err(ArtifactError::UnsafeRoot) => {
+                    report.unsafe_entries += 1;
+                    advance_namespace(cursor);
+                }
+                Err(error) => {
+                    return Err((ReconciliationCursorState::OpenNamespace, error));
+                }
+            }
+        }
+        ReconciliationCursorState::ValidateNamespace(namespace) => {
+            count_traversal_syscall(report);
+            match validate_traversal_dir(&namespace) {
+                Ok(_) => {
+                    cursor.state = ReconciliationCursorState::OpenShard(namespace);
+                }
+                Err(ArtifactError::UnsafeRoot) => {
+                    report.unsafe_entries += 1;
+                    advance_namespace(cursor);
+                }
+                Err(error) => {
+                    return Err((
+                        ReconciliationCursorState::ValidateNamespace(namespace),
+                        error,
+                    ));
+                }
+            }
+        }
+        ReconciliationCursorState::OpenShard(namespace) => {
+            let shard_name = format!("{:02x}", cursor.shard);
+            count_traversal_syscall(report);
+            match open_traversal_dir(namespace.as_raw_fd(), shard_name.as_bytes()) {
+                Ok(shard) => {
+                    cursor.state = ReconciliationCursorState::ValidateShard { namespace, shard };
+                }
+                Err(ArtifactError::NotFound) => advance_reconciliation_shard(cursor, namespace),
+                Err(ArtifactError::UnsafeRoot) => {
+                    report.unsafe_entries += 1;
+                    advance_reconciliation_shard(cursor, namespace);
+                }
+                Err(error) => {
+                    return Err((ReconciliationCursorState::OpenShard(namespace), error));
+                }
+            }
+        }
+        ReconciliationCursorState::ValidateShard { namespace, shard } => {
+            count_traversal_syscall(report);
+            match validate_traversal_dir(&shard) {
+                Ok(stat) => {
+                    cursor.state = read_shard_state(
+                        namespace,
+                        shard,
+                        ShardIdentity(stat.st_dev, stat.st_ino),
+                        vec![0; 8192],
+                        0,
+                        0,
+                        0,
+                    );
+                }
+                Err(ArtifactError::UnsafeRoot) => {
+                    report.unsafe_entries += 1;
+                    advance_reconciliation_shard(cursor, namespace);
+                }
+                Err(error) => {
+                    return Err((
+                        ReconciliationCursorState::ValidateShard { namespace, shard },
+                        error,
+                    ));
+                }
+            }
+        }
+        ReconciliationCursorState::ReadShard {
+            namespace,
+            shard,
+            shard_dev,
+            shard_ino,
+            mut buffer,
+            mut offset,
+            mut filled,
+            mut cookie,
+        } => {
+            if offset == filled {
+                count_traversal_syscall(report);
+                match checkpointed("reconciliation_getdents", || {
+                    getdents64(shard.as_raw_fd(), &mut buffer)
+                }) {
+                    Ok(0) => advance_reconciliation_shard(cursor, namespace),
+                    Ok(count) => {
+                        offset = 0;
+                        filled = count;
+                        cursor.state = read_shard_state(
+                            namespace,
+                            shard,
+                            ShardIdentity(shard_dev, shard_ino),
+                            buffer,
+                            offset,
+                            filled,
+                            cookie,
+                        );
+                    }
+                    Err(error) => {
+                        return Err((
+                            ReconciliationCursorState::SeekShard {
+                                namespace,
+                                shard,
+                                shard_dev,
+                                shard_ino,
+                                buffer,
+                                cookie,
+                            },
+                            error,
+                        ));
+                    }
+                }
+                return Ok(());
+            }
+
+            let (name, after, after_cookie) = match parse_dirent(&buffer[..filled], offset) {
+                Ok(entry) => entry,
+                Err(error) => {
+                    return Err((
+                        read_shard_state(
+                            namespace,
+                            shard,
+                            ShardIdentity(shard_dev, shard_ino),
+                            buffer,
+                            offset,
+                            filled,
+                            cookie,
+                        ),
+                        error,
+                    ));
+                }
+            };
+            if name == b"." || name == b".." {
+                offset = after;
+                cookie = after_cookie;
+                cursor.state = read_shard_state(
+                    namespace,
+                    shard,
+                    ShardIdentity(shard_dev, shard_ino),
+                    buffer,
+                    offset,
+                    filled,
+                    cookie,
+                );
+                return Ok(());
+            }
+            let shard_name = format!("{:02x}", cursor.shard);
+            let Ok(name) = String::from_utf8(name) else {
+                report.scanned += 1;
+                report.malformed += 1;
+                offset = after;
+                cookie = after_cookie;
+                cursor.state = read_shard_state(
+                    namespace,
+                    shard,
+                    ShardIdentity(shard_dev, shard_ino),
+                    buffer,
+                    offset,
+                    filled,
+                    cookie,
+                );
+                return Ok(());
+            };
+            if !valid_final_name(&name, &shard_name) {
+                report.scanned += 1;
+                if cursor.namespace != FINAL_NAMESPACE || !valid_temp_name(&name) {
+                    report.malformed += 1;
+                }
+                offset = after;
+                cookie = after_cookie;
+                cursor.state = read_shard_state(
+                    namespace,
+                    shard,
+                    ShardIdentity(shard_dev, shard_ino),
+                    buffer,
+                    offset,
+                    filled,
+                    cookie,
+                );
+                return Ok(());
+            }
+
+            #[cfg(debug_assertions)]
+            let _ = crate::test_support::checkpoint("reconciliation_before_stat");
+            count_traversal_syscall(report);
+            let stat = match checkpointed("reconciliation_stat", || {
+                record_traversal_call();
+                nofollow_stat(shard.as_raw_fd(), &name)
+            }) {
+                Ok(stat) => Some(stat),
+                Err(ArtifactError::NotFound) => None,
+                Err(ArtifactError::UnsafeRoot) => {
+                    report.unsafe_entries += 1;
+                    None
+                }
+                Err(error) => {
+                    return Err((
+                        read_shard_state(
+                            namespace,
+                            shard,
+                            ShardIdentity(shard_dev, shard_ino),
+                            buffer,
+                            offset,
+                            filled,
+                            cookie,
+                        ),
+                        error,
+                    ));
+                }
+            };
+            report.scanned += 1;
+            offset = after;
+            cookie = after_cookie;
+            if let Some(stat) = stat {
+                if check_file(&stat).is_err() {
+                    report.unsafe_entries += 1;
+                } else {
+                    let candidate_namespace = if cursor.namespace == FINAL_NAMESPACE {
+                        report.final_entries += 1;
+                        ReconciliationNamespace::Final
+                    } else {
+                        report.quarantined_entries += 1;
+                        ReconciliationNamespace::Quarantine
+                    };
+                    entries.push(ReconciliationCandidate {
+                        root_dev: root.dev,
+                        root_ino: root.ino,
+                        namespace: candidate_namespace,
+                        shard: shard_name,
+                        shard_dev,
+                        shard_ino,
+                        name,
+                        dev: stat.st_dev,
+                        ino: stat.st_ino,
+                        modified_seconds: stat.st_mtime,
+                        modified_nanoseconds: stat.st_mtime_nsec,
+                    });
+                }
+            }
+            cursor.state = read_shard_state(
+                namespace,
+                shard,
+                ShardIdentity(shard_dev, shard_ino),
+                buffer,
+                offset,
+                filled,
+                cookie,
+            );
+        }
+        ReconciliationCursorState::SeekShard {
+            namespace,
+            shard,
+            shard_dev,
+            shard_ino,
+            buffer,
+            cookie,
+        } => {
+            count_traversal_syscall(report);
+            record_traversal_call();
+            if unsafe { libc::lseek(shard.as_raw_fd(), cookie, libc::SEEK_SET) } != cookie {
+                return Err((
+                    ReconciliationCursorState::SeekShard {
+                        namespace,
+                        shard,
+                        shard_dev,
+                        shard_ino,
+                        buffer,
+                        cookie,
+                    },
+                    ArtifactError::Storage,
+                ));
+            }
+            cursor.state = read_shard_state(
+                namespace,
+                shard,
+                ShardIdentity(shard_dev, shard_ino),
+                buffer,
+                0,
+                0,
+                cookie,
+            );
+        }
+        ReconciliationCursorState::Transition => unreachable!("temporary scanner state"),
+    }
+    Ok(())
+}
+
+fn count_traversal_syscall(report: &mut ReconciliationScan) {
+    report.traversal_syscalls += 1;
+}
+
+fn read_shard_state(
+    namespace: OwnedFd,
+    shard: OwnedFd,
+    shard_identity: ShardIdentity,
+    buffer: Vec<u8>,
+    offset: usize,
+    filled: usize,
+    cookie: i64,
+) -> ReconciliationCursorState {
+    let ShardIdentity(shard_dev, shard_ino) = shard_identity;
+    ReconciliationCursorState::ReadShard {
         namespace,
         shard,
-        position,
+        shard_dev,
+        shard_ino,
+        buffer,
+        offset,
+        filled,
+        cookie,
     }
 }
 
+fn open_traversal_dir(parent: i32, name: &[u8]) -> Result<OwnedFd, ArtifactError> {
+    let name = crate::store::c_name(name)?;
+    record_traversal_call();
+    let raw = unsafe {
+        libc::openat(
+            parent,
+            name.as_ptr(),
+            libc::O_RDONLY | libc::O_DIRECTORY | libc::O_CLOEXEC | libc::O_NOFOLLOW,
+        )
+    };
+    if raw < 0 {
+        match unsafe { *libc::__errno_location() } {
+            libc::ENOENT => Err(ArtifactError::NotFound),
+            libc::EACCES | libc::ELOOP | libc::ENOTDIR => Err(ArtifactError::UnsafeRoot),
+            _ => Err(ArtifactError::Storage),
+        }
+    } else {
+        Ok(unsafe { OwnedFd::from_raw_fd(raw) })
+    }
+}
+
+fn validate_traversal_dir(directory: &OwnedFd) -> Result<libc::stat, ArtifactError> {
+    record_traversal_call();
+    let stat = stat_fd(directory.as_raw_fd())?;
+    if stat.st_uid != unsafe { libc::geteuid() }
+        || (stat.st_mode & libc::S_IFMT) != libc::S_IFDIR
+        || (stat.st_mode & 0o777) != 0o700
+    {
+        Err(ArtifactError::UnsafeRoot)
+    } else {
+        Ok(stat)
+    }
+}
+
+fn getdents64(directory: i32, buffer: &mut [u8]) -> Result<usize, ArtifactError> {
+    record_traversal_call();
+    let count = unsafe {
+        libc::syscall(
+            libc::SYS_getdents64,
+            directory,
+            buffer.as_mut_ptr(),
+            buffer.len(),
+        )
+    };
+    if count < 0 {
+        Err(ArtifactError::Storage)
+    } else {
+        usize::try_from(count).map_err(|_| ArtifactError::Storage)
+    }
+}
+
+fn parse_dirent(buffer: &[u8], offset: usize) -> Result<(Vec<u8>, usize, i64), ArtifactError> {
+    const HEADER: usize = 19;
+    if buffer.len().saturating_sub(offset) < HEADER {
+        return Err(ArtifactError::Storage);
+    }
+    let record_length = usize::from(u16::from_ne_bytes([
+        buffer[offset + 16],
+        buffer[offset + 17],
+    ]));
+    let after = offset
+        .checked_add(record_length)
+        .filter(|after| record_length >= HEADER && *after <= buffer.len())
+        .ok_or(ArtifactError::Storage)?;
+    let raw_name = &buffer[offset + HEADER..after];
+    let end = raw_name
+        .iter()
+        .position(|byte| *byte == 0)
+        .ok_or(ArtifactError::Storage)?;
+    let cookie = i64::from_ne_bytes(
+        buffer[offset + 8..offset + 16]
+            .try_into()
+            .map_err(|_| ArtifactError::Storage)?,
+    );
+    Ok((raw_name[..end].to_vec(), after, cookie))
+}
+
 fn namespace_name(namespace: u8) -> &'static [u8] {
     match namespace {
         FINAL_NAMESPACE => b"sha256",
@@ -477,22 +873,19 @@ fn namespace_name(namespace: u8) -> &'static [u8] {
     }
 }
 
-fn advance_shard(cursor: &mut ReconciliationCursor) {
+fn advance_reconciliation_shard(cursor: &mut ReconciliationCursor, namespace: OwnedFd) {
-    debug_assert_ne!(cursor.shard, u8::MAX);
+    if cursor.shard == u8::MAX {
-    cursor.shard += 1;
+        advance_namespace(cursor);
-    cursor.position = 0;
+    } else {
+        cursor.shard += 1;
+        cursor.state = ReconciliationCursorState::OpenShard(namespace);
+    }
 }
 
 fn advance_namespace(cursor: &mut ReconciliationCursor) {
     cursor.namespace = cursor.namespace.saturating_add(1);
     cursor.shard = 0;
-    cursor.position = 0;
+    cursor.state = ReconciliationCursorState::OpenNamespace;
-}
-
-fn consume_scan_work(report: &mut ReconciliationScan, remaining: &mut usize) {
-    debug_assert!(*remaining > 0);
-    *remaining -= 1;
-    report.scanned += 1;
 }
 
 fn ensure_reconciliation_root(
@@ -530,10 +923,8 @@ fn revalidate_candidate(
     check_file(&stat).map_err(|_| ArtifactError::UnsafeRoot)
 }
 
-/// Recovery follows a successful cross-directory rename. The candidate remains
+/// After cross-directory rename, only canonical name and inode remain common;
-/// bound to the original final shard, so the quarantine shard identity cannot
+/// the quarantine shard cannot equal the capability's recorded final shard.
-/// equal its recorded shard identity; only the canonical name and inode are
-/// revalidated here.
 fn revalidate_moved_candidate(
     candidate: &ReconciliationCandidate,
     shard: i32,
@@ -561,9 +952,7 @@ fn fsync_mutation(
     }
 }
 
-/// Persist both directory entry changes even if the first fsync reports an
+/// Attempts both directory fsyncs; either failure leaves a retryable ambiguity.
-/// error.  Either failure remains an ambiguity, but skipping the second fsync
-/// would needlessly enlarge the crash window.
 fn fsync_quarantine_dirs(
     final_shard: &std::os::fd::OwnedFd,
     quarantine_shard: &std::os::fd::OwnedFd,
@@ -581,132 +970,6 @@ fn valid_final_name(name: &str, shard: &str) -> bool {
     name.len() == 64 && name.starts_with(shard) && name.bytes().all(is_lower_hex)
 }
 
-struct DirectoryStream {
-    directory: *mut libc::DIR,
-}
-
-impl DirectoryStream {
-    fn open(parent: i32, position: i64) -> Result<Self, ArtifactError> {
-        let duplicate = crate::store::duplicate_fd(parent)?;
-        let raw = std::os::fd::IntoRawFd::into_raw_fd(duplicate);
-        let directory = unsafe { libc::fdopendir(raw) };
-        if directory.is_null() {
-            unsafe {
-                libc::close(raw);
-            }
-            return Err(ArtifactError::Storage);
-        }
-        if position != 0 {
-            unsafe {
-                libc::seekdir(directory, position as libc::c_long);
-            }
-        }
-        Ok(Self { directory })
-    }
-
-    fn next(&mut self) -> Result<Option<(Vec<u8>, i64)>, ArtifactError> {
-        loop {
-            unsafe {
-                *libc::__errno_location() = 0;
-            }
-            let entry = unsafe { libc::readdir(self.directory) };
-            if entry.is_null() {
-                if unsafe { *libc::__errno_location() } != 0 {
-                    return Err(ArtifactError::Storage);
-                }
-                return Ok(None);
-            }
-            let after = unsafe { libc::telldir(self.directory) } as i64;
-            let name = unsafe { std::ffi::CStr::from_ptr((*entry).d_name.as_ptr()) }.to_bytes();
-            if name == b"." || name == b".." {
-                continue;
-            }
-            return Ok(Some((name.to_vec(), after)));
-        }
-    }
-}
-
-impl Drop for DirectoryStream {
-    fn drop(&mut self) {
-        unsafe {
-            libc::closedir(self.directory);
-        }
-    }
-}
-
-fn list_names(
-    parent: i32,
-    remaining: &mut usize,
-    report: &mut TempScan,
-) -> Result<Vec<String>, ArtifactError> {
-    let dot = std::ffi::CString::new(".").map_err(|_| ArtifactError::Storage)?;
-    let raw = unsafe {
-        libc::openat(
-            parent,
-            dot.as_ptr(),
-            libc::O_RDONLY | libc::O_DIRECTORY | libc::O_CLOEXEC,
-        )
-    };
-    if raw < 0 {
-        return Err(ArtifactError::Storage);
-    }
-    let directory = unsafe { libc::fdopendir(raw) };
-    if directory.is_null() {
-        unsafe {
-            libc::close(raw);
-        }
-        return Err(ArtifactError::Storage);
-    }
-    let mut names = Vec::new();
-    loop {
-        if *remaining == 0 {
-            break;
-        }
-        unsafe {
-            *libc::__errno_location() = 0;
-        }
-        let entry = unsafe { libc::readdir(directory) };
-        if entry.is_null() {
-            if unsafe { *libc::__errno_location() } != 0 {
-                unsafe {
-                    libc::closedir(directory);
-                }
-                return Err(ArtifactError::Storage);
-            }
-            break;
-        }
-        let name = unsafe { std::ffi::CStr::from_ptr((*entry).d_name.as_ptr()) }.to_bytes();
-        if name == b"." || name == b".." {
-            continue;
-        }
-        report.scanned += 1;
-        *remaining -= 1;
-        if let Ok(name) = std::str::from_utf8(name) {
-            names.push(name.to_owned());
-        }
-    }
-    unsafe {
-        libc::closedir(directory);
-    }
-    Ok(names)
-}
-fn valid_temp_name(name: &str) -> bool {
-    let Some(rest) = name.strip_prefix(ArtifactStore::temp_prefix()) else {
-        return false;
-    };
-    let mut fields = rest.split('-');
-    let (Some(nonce), Some(pid), Some(sequence), None) =
-        (fields.next(), fields.next(), fields.next(), fields.next())
-    else {
-        return false;
-    };
-    nonce.len() == 32
-        && nonce.bytes().all(is_lower_hex)
-        && !pid.is_empty()
-        && pid.bytes().all(|byte| byte.is_ascii_digit())
-        && !sequence.is_empty()
-        && sequence.bytes().all(|byte| byte.is_ascii_digit())
-}
 fn is_lower_hex(byte: u8) -> bool {
     byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte)
 }
@@ -719,3 +982,8 @@ fn nofollow_stat(parent: i32, name: &str) -> Result<libc::stat, ArtifactError> {
         Err(crate::store::classify_errno())
     }
 }
+
+fn record_traversal_call() {
+    #[cfg(debug_assertions)]
+    crate::test_support::record_traversal_call();
+}

crates/crank-artifacts/src/lib.rs

@@ -12,6 +12,7 @@ mod housekeeping;
 mod legacy;
 mod model;
 mod store;
+mod temp_scan;
 
 #[cfg(debug_assertions)]
 #[doc(hidden)]
@@ -21,6 +22,7 @@ pub use error::ArtifactError;
 pub use housekeeping::{ReconciliationCandidate, ReconciliationCursor, StaleTemp, TempScan};
 pub use model::{
     ArtifactRef, MAX_ARTIFACT_BYTES, MAX_SOURCE_BYTES, ReconciliationMutation,
-    ReconciliationNamespace, ReconciliationScan, RegisteredArtifact, StoredArtifact,
+    ReconciliationNamespace, ReconciliationScan, ReconciliationScanStop, RegisteredArtifact,
+    StoredArtifact,
 };
 pub use store::ArtifactStore;

crates/crank-artifacts/src/model.rs

@@ -85,17 +85,32 @@ pub enum ReconciliationNamespace {
     Quarantine,
 }
 
+/// Why a bounded reconciliation page stopped.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ReconciliationScanStop {
+    Complete,
+    ScanBudget,
+    ResultLimit,
+    Retryable,
+}
+
 /// Bounded, redacted progress report from a reconciliation page.
 ///
 /// `continuation` is opaque evidence for resuming an unchanged namespace. It
 /// intentionally cannot be created from a caller-supplied path or digest.
 #[derive(Debug)]
 pub struct ReconciliationScan {
+    /// Non-dot directory entries classified by this page.
     pub scanned: usize,
+    /// Traversal syscall budget consumed by this page. It equals syscall
+    /// attempts in production; debug fault injection may stop an attempt just
+    /// before the syscall. Root validation and locking are fixed overhead.
+    pub traversal_syscalls: usize,
     pub malformed: usize,
     pub unsafe_entries: usize,
     pub final_entries: usize,
     pub quarantined_entries: usize,
+    pub stop: ReconciliationScanStop,
     pub continuation: Option<crate::ReconciliationCursor>,
 }
 
@@ -103,10 +118,12 @@ impl ReconciliationScan {
     pub(crate) fn empty(continuation: Option<crate::ReconciliationCursor>) -> Self {
         Self {
             scanned: 0,
+            traversal_syscalls: 0,
             malformed: 0,
             unsafe_entries: 0,
             final_entries: 0,
             quarantined_entries: 0,
+            stop: ReconciliationScanStop::Complete,
             continuation,
         }
     }

crates/crank-artifacts/src/temp_scan.rs

@@ -0,0 +1,78 @@
+use crate::{ArtifactError, ArtifactStore, TempScan};
+
+pub(crate) fn valid_temp_name(name: &str) -> bool {
+    let Some(rest) = name.strip_prefix(ArtifactStore::temp_prefix()) else {
+        return false;
+    };
+    let mut fields = rest.split('-');
+    let (Some(nonce), Some(pid), Some(sequence), None) =
+        (fields.next(), fields.next(), fields.next(), fields.next())
+    else {
+        return false;
+    };
+    nonce.len() == 32
+        && nonce
+            .bytes()
+            .all(|byte| byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte))
+        && !pid.is_empty()
+        && pid.bytes().all(|byte| byte.is_ascii_digit())
+        && !sequence.is_empty()
+        && sequence.bytes().all(|byte| byte.is_ascii_digit())
+}
+
+pub(crate) fn list_names(
+    parent: i32,
+    remaining: &mut usize,
+    report: &mut TempScan,
+) -> Result<Vec<String>, ArtifactError> {
+    let dot = std::ffi::CString::new(".").map_err(|_| ArtifactError::Storage)?;
+    let raw = unsafe {
+        libc::openat(
+            parent,
+            dot.as_ptr(),
+            libc::O_RDONLY | libc::O_DIRECTORY | libc::O_CLOEXEC,
+        )
+    };
+    if raw < 0 {
+        return Err(ArtifactError::Storage);
+    }
+    let directory = unsafe { libc::fdopendir(raw) };
+    if directory.is_null() {
+        unsafe {
+            libc::close(raw);
+        }
+        return Err(ArtifactError::Storage);
+    }
+    let mut names = Vec::new();
+    loop {
+        if *remaining == 0 {
+            break;
+        }
+        unsafe {
+            *libc::__errno_location() = 0;
+        }
+        let entry = unsafe { libc::readdir(directory) };
+        if entry.is_null() {
+            if unsafe { *libc::__errno_location() } != 0 {
+                unsafe {
+                    libc::closedir(directory);
+                }
+                return Err(ArtifactError::Storage);
+            }
+            break;
+        }
+        let name = unsafe { std::ffi::CStr::from_ptr((*entry).d_name.as_ptr()) }.to_bytes();
+        if name == b"." || name == b".." {
+            continue;
+        }
+        report.scanned += 1;
+        *remaining -= 1;
+        if let Ok(name) = std::str::from_utf8(name) {
+            names.push(name.to_owned());
+        }
+    }
+    unsafe {
+        libc::closedir(directory);
+    }
+    Ok(names)
+}

crates/crank-artifacts/src/test_support.rs

@@ -1,10 +1,28 @@
 //! Test-only crash and fault control for syscall-coupled storage checkpoints.
 
 use std::{
-    sync::{Condvar, Mutex, OnceLock},
+    collections::BTreeMap,
+    sync::{
+        Condvar, Mutex, OnceLock,
+        atomic::{AtomicUsize, Ordering},
+    },
     time::Duration,
 };
 
+static TRAVERSAL_CALLS: AtomicUsize = AtomicUsize::new(0);
+
+pub fn reset_traversal_calls() {
+    TRAVERSAL_CALLS.store(0, Ordering::SeqCst);
+}
+
+pub fn traversal_calls() -> usize {
+    TRAVERSAL_CALLS.load(Ordering::SeqCst)
+}
+
+pub(crate) fn record_traversal_call() {
+    TRAVERSAL_CALLS.fetch_add(1, Ordering::SeqCst);
+}
+
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub enum FaultAction {
     Exit,
@@ -13,7 +31,8 @@ pub enum FaultAction {
 }
 
 struct State {
-    checkpoint: Option<(String, FaultAction)>,
+    checkpoint: Option<(String, FaultAction, usize)>,
+    hits: BTreeMap<String, usize>,
     held: bool,
 }
 fn slot() -> &'static (Mutex<State>, Condvar) {
@@ -22,6 +41,7 @@ fn slot() -> &'static (Mutex<State>, Condvar) {
         (
             Mutex::new(State {
                 checkpoint: None,
+                hits: BTreeMap::new(),
                 held: false,
             }),
             Condvar::new(),
@@ -30,10 +50,25 @@ fn slot() -> &'static (Mutex<State>, Condvar) {
 }
 
 pub fn set_checkpoint(stage: impl Into<String>, action: FaultAction) {
+    set_checkpoint_on_hit(stage, action, 1);
+}
+
+pub fn set_checkpoint_on_hit(stage: impl Into<String>, action: FaultAction, hit: usize) {
+    assert!(hit > 0, "checkpoint hit is one-based");
+    let (lock, _) = slot();
+    let mut state = lock.lock().expect("artifact test checkpoint lock poisoned");
+    state.checkpoint = Some((stage.into(), action, hit));
+    state.hits.clear();
+}
+
+pub fn checkpoint_hits(stage: &str) -> usize {
     let (lock, _) = slot();
     lock.lock()
         .expect("artifact test checkpoint lock poisoned")
-        .checkpoint = Some((stage.into(), action));
+        .hits
+        .get(stage)
+        .copied()
+        .unwrap_or(0)
 }
 
 pub fn clear_checkpoint() {
@@ -55,10 +90,17 @@ pub fn wait_until_held(timeout: Duration) -> bool {
 pub(crate) fn checkpoint(stage: &str) -> Option<FaultAction> {
     let (lock, wake) = slot();
     let mut state = lock.lock().expect("artifact test checkpoint lock poisoned");
-    let action = state
+    let configured = state.checkpoint.clone();
-        .checkpoint
+    let action = configured.and_then(|(expected, action, target_hit)| {
-        .as_ref()
+        if expected != stage {
-        .and_then(|(expected, action)| (expected == stage).then_some(*action));
+            return None;
+        }
+        let hit = state.hits.entry(stage.to_owned()).or_insert(0);
+        if *hit == 0 {
+            *hit = 1;
+        }
+        (*hit == target_hit).then_some(action)
+    });
     if action == Some(FaultAction::Hold) {
         state.held = true;
         wake.notify_all();
@@ -73,9 +115,13 @@ pub(crate) fn checkpoint(stage: &str) -> Option<FaultAction> {
 
 pub(crate) fn action(stage: &str) -> Option<FaultAction> {
     let (lock, _) = slot();
-    let state = lock.lock().expect("artifact test checkpoint lock poisoned");
+    let mut state = lock.lock().expect("artifact test checkpoint lock poisoned");
+    *state.hits.entry(stage.to_owned()).or_insert(0) += 1;
+    let hit = state.hits[stage];
     state
         .checkpoint
         .as_ref()
-        .and_then(|(expected, action)| (expected == stage).then_some(*action))
+        .and_then(|(expected, action, target_hit)| {
+            (expected == stage && hit == *target_hit).then_some(*action)
+        })
 }

crates/crank-artifacts/tests/reconciliation.rs

@@ -1,28 +1,28 @@
 use std::{
     fs,
-    os::unix::fs::PermissionsExt,
+    os::unix::fs::{MetadataExt, PermissionsExt},
     path::PathBuf,
-    process::Command,
+    sync::atomic::{AtomicU64, Ordering},
-    sync::{
-        Arc,
-        atomic::{AtomicU64, Ordering},
-    },
-    thread,
-    time::Duration,
 };
 
+#[cfg(debug_assertions)]
+use std::{process::Command, sync::Arc, thread, time::Duration};
+
 use crank_artifacts::{
     ArtifactError, ArtifactStore, ReconciliationMutation, ReconciliationNamespace,
+    ReconciliationScanStop,
 };
 
 #[cfg(debug_assertions)]
 use crank_artifacts::test_support::{
-    FaultAction, clear_checkpoint, set_checkpoint, wait_until_held,
+    FaultAction, checkpoint_hits, clear_checkpoint, reset_traversal_calls, set_checkpoint,
+    set_checkpoint_on_hit, traversal_calls, wait_until_held,
 };
 #[cfg(debug_assertions)]
 use std::sync::{Mutex, OnceLock};
 
 static NEXT_ROOT: AtomicU64 = AtomicU64::new(0);
+const FULL_SCAN_BUDGET: usize = 2048;
 
 struct TestRoot(PathBuf);
 
@@ -61,16 +61,31 @@ fn paginates_final_entries_without_disclosing_locations() {
 
     let mut continuation = None;
     let mut candidates = Vec::new();
+    let mut final_entries = 0;
+    let mut scanned = 0;
     loop {
         let (report, mut page) = store.scan_reconciliation(continuation, 1, 1).unwrap();
+        assert_eq!(report.traversal_syscalls, 1);
         assert!(report.scanned <= 1);
+        assert_eq!(report.final_entries, page.len());
+        assert_eq!(report.quarantined_entries, 0);
         assert!(page.len() <= 1);
+        scanned += report.scanned;
+        final_entries += report.final_entries;
+        if let Some(cursor) = report.continuation.as_ref() {
+            assert_eq!(format!("{cursor:?}"), "ReconciliationCursor(..)");
+        }
+        for candidate in &page {
+            assert_eq!(format!("{candidate:?}"), "ReconciliationCandidate(..)");
+        }
         candidates.append(&mut page);
         continuation = report.continuation;
         if continuation.is_none() {
             break;
         }
     }
+    assert_eq!(scanned, 2);
+    assert_eq!(final_entries, 2);
     assert_eq!(candidates.len(), 2);
     assert!(
         candidates
@@ -83,8 +98,14 @@ fn paginates_final_entries_without_disclosing_locations() {
             ReconciliationMutation::Quarantined
         );
     }
-    let (_, quarantined_entries) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (report, quarantined_entries) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     assert_eq!(quarantined_entries.len(), 2);
+    assert_eq!(report.scanned, 2);
+    assert_eq!(report.final_entries, 0);
+    assert_eq!(report.quarantined_entries, 2);
+    assert_eq!(report.stop, ReconciliationScanStop::Complete);
     assert!(
         quarantined_entries
             .iter()
@@ -121,13 +142,98 @@ fn classifies_malformed_and_unsafe_entries_with_a_usable_page() {
     );
     fs::hard_link(shard.join("not-a-digest"), shard.join(hardlink_name)).unwrap();
 
-    let (report, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (report, candidates) = store
-    assert!(report.malformed >= 1);
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
-    assert!(report.unsafe_entries >= 3);
+        .unwrap();
+    assert_eq!(report.scanned, 5);
+    assert_eq!(report.malformed, 1);
+    assert_eq!(report.unsafe_entries, 3);
+    assert_eq!(report.final_entries, 1);
+    assert_eq!(report.quarantined_entries, 0);
     assert_eq!(candidates.len(), 1);
     assert!(report.continuation.is_none());
 }
 
+#[test]
+fn skips_a_valid_publish_temp_without_classifying_it_as_malformed() {
+    let root = TestRoot::new("valid-temp");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let stored = store.put(b"reconciliation with concurrent temp").unwrap();
+    let shard = root
+        .0
+        .join("sha256")
+        .join(&stored.artifact_ref.digest_hex()[..2]);
+    fs::write(
+        shard.join(".crank-artifact-tmp-v1-0123456789abcdef0123456789abcdef-7-9"),
+        b"in-flight",
+    )
+    .unwrap();
+
+    let (report, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.scanned, 2);
+    assert_eq!(report.malformed, 0);
+    assert_eq!(report.final_entries, 1);
+    assert_eq!(candidates.len(), 1);
+}
+
+#[test]
+fn unsafe_canonical_shard_does_not_starve_later_shards() {
+    let root = TestRoot::new("unsafe-shard");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let sha = root.0.join("sha256");
+    fs::create_dir(&sha).unwrap();
+    fs::set_permissions(&sha, fs::Permissions::from_mode(0o700)).unwrap();
+    let unsafe_shard = sha.join("00");
+    fs::create_dir(&unsafe_shard).unwrap();
+    fs::set_permissions(&unsafe_shard, fs::Permissions::from_mode(0o000)).unwrap();
+    let shard = sha.join("ff");
+    fs::create_dir(&shard).unwrap();
+    fs::set_permissions(&shard, fs::Permissions::from_mode(0o700)).unwrap();
+    let name = "f".repeat(64);
+    fs::write(shard.join(&name), b"later valid entry").unwrap();
+    fs::set_permissions(shard.join(&name), fs::Permissions::from_mode(0o400)).unwrap();
+
+    let (report, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::Complete);
+    assert_eq!(report.unsafe_entries, 1);
+    assert_eq!(report.final_entries, 1);
+    assert!(candidates.iter().any(|candidate| {
+        candidate.namespace() == ReconciliationNamespace::Final
+            && format!("{candidate:?}") == "ReconciliationCandidate(..)"
+    }));
+    fs::set_permissions(unsafe_shard, fs::Permissions::from_mode(0o700)).unwrap();
+}
+
+#[test]
+#[cfg(debug_assertions)]
+fn traversal_never_exceeds_the_exact_syscall_budget() {
+    let root = TestRoot::new("syscall-budget");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    store.put(b"bounded traversal syscall accounting").unwrap();
+
+    for budget in 0..=8 {
+        reset_traversal_calls();
+        let (report, _) = store.scan_reconciliation(None, budget, 8).unwrap();
+        assert!(report.traversal_syscalls <= budget, "budget={budget}");
+        assert_eq!(report.traversal_syscalls, budget, "budget={budget}");
+        assert_eq!(traversal_calls(), report.traversal_syscalls);
+        assert_eq!(report.stop, ReconciliationScanStop::ScanBudget);
+    }
+
+    reset_traversal_calls();
+    let (report, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 1)
+        .unwrap();
+    assert_eq!(candidates.len(), 1);
+    assert_eq!(report.stop, ReconciliationScanStop::ResultLimit);
+    assert!(report.traversal_syscalls <= FULL_SCAN_BUDGET);
+    assert_eq!(traversal_calls(), report.traversal_syscalls);
+}
+
 #[test]
 fn rejects_a_cursor_from_another_store() {
     let first_root = TestRoot::new("foreign-cursor-first");
@@ -157,7 +263,7 @@ fn disappearing_entry_after_readdir_is_skipped() {
         .join(stored.artifact_ref.digest_hex());
     set_checkpoint("reconciliation_before_stat", FaultAction::Hold);
     let scan_store = Arc::clone(&store);
-    let scan = thread::spawn(move || scan_store.scan_reconciliation(None, 512, 8));
+    let scan = thread::spawn(move || scan_store.scan_reconciliation(None, FULL_SCAN_BUDGET, 8));
     assert!(wait_until_held(Duration::from_secs(2)));
     fs::set_permissions(&path, fs::Permissions::from_mode(0o600)).unwrap();
     fs::remove_file(&path).unwrap();
@@ -172,7 +278,9 @@ fn quarantine_and_delete_are_idempotent() {
     let root = TestRoot::new("mutation");
     let store = ArtifactStore::open(&root.0).unwrap();
     let stored = store.put(b"reconciliation mutation").unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let candidate = candidates.into_iter().next().unwrap();
 
     assert_eq!(
@@ -188,7 +296,9 @@ fn quarantine_and_delete_are_idempotent() {
         ReconciliationMutation::AlreadyQuarantined
     );
 
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let quarantined = candidates
         .into_iter()
         .find(|candidate| candidate.namespace() == ReconciliationNamespace::Quarantine)
@@ -212,7 +322,9 @@ fn quarantine_preserves_old_inode_when_the_digest_is_republished() {
     let root = TestRoot::new("no-clobber");
     let store = ArtifactStore::open(&root.0).unwrap();
     let stored = store.put(b"reconciliation no-clobber").unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let original = candidates.into_iter().next().unwrap();
     assert_eq!(
         store.quarantine_reconciliation(original.clone()).unwrap(),
@@ -227,7 +339,9 @@ fn quarantine_preserves_old_inode_when_the_digest_is_republished() {
         store.read(&stored.artifact_ref).unwrap(),
         b"reconciliation no-clobber"
     );
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     assert_eq!(candidates.len(), 2);
     assert!(
         candidates
@@ -241,6 +355,42 @@ fn quarantine_preserves_old_inode_when_the_digest_is_republished() {
     );
 }
 
+#[test]
+fn quarantine_refuses_to_replace_an_unrelated_inode() {
+    let root = TestRoot::new("unrelated-collision");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let stored = store.put(b"canonical final survives collision").unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    let candidate = candidates.into_iter().next().unwrap();
+    let digest = stored.artifact_ref.digest_hex();
+    let quarantine = root.0.join("quarantine");
+    let shard = quarantine.join(&digest[..2]);
+    fs::create_dir(&quarantine).unwrap();
+    fs::create_dir(&shard).unwrap();
+    fs::set_permissions(&quarantine, fs::Permissions::from_mode(0o700)).unwrap();
+    fs::set_permissions(&shard, fs::Permissions::from_mode(0o700)).unwrap();
+    let collision = shard.join(digest);
+    fs::write(&collision, b"unrelated quarantine inode").unwrap();
+    fs::set_permissions(&collision, fs::Permissions::from_mode(0o400)).unwrap();
+    let collision_ino = fs::metadata(&collision).unwrap().ino();
+
+    assert_eq!(
+        store.quarantine_reconciliation(candidate),
+        Err(ArtifactError::UnsafeRoot)
+    );
+    assert_eq!(
+        store.read(&stored.artifact_ref).unwrap(),
+        b"canonical final survives collision"
+    );
+    assert_eq!(fs::read(collision).unwrap(), b"unrelated quarantine inode");
+    assert_eq!(
+        fs::metadata(shard.join(digest)).unwrap().ino(),
+        collision_ino
+    );
+}
+
 #[test]
 fn delete_fails_closed_for_a_replaced_or_hardlinked_quarantined_inode() {
     let root = TestRoot::new("quarantine-replaced");
@@ -248,11 +398,15 @@ fn delete_fails_closed_for_a_replaced_or_hardlinked_quarantined_inode() {
     let stored = store
         .put(b"reconciliation quarantined replacement")
         .unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     store
         .quarantine_reconciliation(candidates.into_iter().next().unwrap())
         .unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let candidate = candidates.into_iter().next().unwrap();
     let path = root
         .0
@@ -274,6 +428,40 @@ fn delete_fails_closed_for_a_replaced_or_hardlinked_quarantined_inode() {
     assert!(sibling.exists());
 }
 
+#[test]
+fn delete_revalidates_a_post_scan_hardlink_without_inode_replacement() {
+    let root = TestRoot::new("quarantine-hardlink");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    let stored = store.put(b"same quarantined inode gains a link").unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    store
+        .quarantine_reconciliation(candidates.into_iter().next().unwrap())
+        .unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    let candidate = candidates.into_iter().next().unwrap();
+    let path = root
+        .0
+        .join("quarantine")
+        .join(&stored.artifact_ref.digest_hex()[..2])
+        .join(stored.artifact_ref.digest_hex());
+    let sibling = path.with_extension("link");
+    let original_ino = fs::metadata(&path).unwrap().ino();
+    fs::hard_link(&path, &sibling).unwrap();
+
+    assert_eq!(
+        store.delete_quarantined_reconciliation(candidate),
+        Err(ArtifactError::UnsafeRoot)
+    );
+    assert_eq!(fs::metadata(&path).unwrap().nlink(), 2);
+    assert_eq!(fs::metadata(&path).unwrap().ino(), original_ino);
+    assert!(path.exists());
+    assert!(sibling.exists());
+}
+
 #[test]
 #[cfg(debug_assertions)]
 fn delete_fsync_ambiguity_is_retryable_and_recovers_as_absent() {
@@ -281,11 +469,15 @@ fn delete_fsync_ambiguity_is_retryable_and_recovers_as_absent() {
     let root = TestRoot::new("delete-fsync");
     let store = ArtifactStore::open(&root.0).unwrap();
     store.put(b"reconciliation delete fsync").unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     store
         .quarantine_reconciliation(candidates.into_iter().next().unwrap())
         .unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let candidate = candidates.into_iter().next().unwrap();
     set_checkpoint("reconciliation_delete_fsync", FaultAction::Fail);
     assert_eq!(
@@ -306,7 +498,9 @@ fn rejects_a_replaced_final_inode() {
     let root = TestRoot::new("replaced");
     let store = ArtifactStore::open(&root.0).unwrap();
     let stored = store.put(b"reconciliation original").unwrap();
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    let (_, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let candidate = candidates.into_iter().next().unwrap();
     let path = root
         .0
@@ -328,36 +522,175 @@ fn rejects_a_replaced_final_inode() {
 #[cfg(debug_assertions)]
 fn retries_after_post_rename_fsync_ambiguity() {
     let _guard = fault_guard();
-    let root = TestRoot::new("fsync");
+    for stage in [
-    let store = ArtifactStore::open(&root.0).unwrap();
+        "reconciliation_quarantine_source_fsync",
-    store.put(b"reconciliation fault").unwrap();
+        "reconciliation_quarantine_destination_fsync",
-    let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+    ] {
+        let root = TestRoot::new(stage);
+        let store = ArtifactStore::open(&root.0).unwrap();
+        store.put(stage.as_bytes()).unwrap();
+        let (_, candidates) = store
+            .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+            .unwrap();
+        let candidate = candidates.into_iter().next().unwrap();
+        set_checkpoint(stage, FaultAction::Fail);
+        assert_eq!(
+            store.quarantine_reconciliation(candidate.clone()).unwrap(),
+            ReconciliationMutation::Retryable,
+            "stage={stage}"
+        );
+        assert_eq!(
+            checkpoint_hits("reconciliation_quarantine_source_fsync"),
+            1,
+            "stage={stage}"
+        );
+        assert_eq!(
+            checkpoint_hits("reconciliation_quarantine_destination_fsync"),
+            1,
+            "stage={stage}"
+        );
+        clear_checkpoint();
+        assert_eq!(
+            store.quarantine_reconciliation(candidate).unwrap(),
+            ReconciliationMutation::AlreadyQuarantined,
+            "stage={stage}"
+        );
+    }
+}
+
+#[test]
+#[cfg(debug_assertions)]
+fn mutation_syscall_failures_preserve_retryable_state() {
+    let _guard = fault_guard();
+
+    let rename_root = TestRoot::new("rename-fail");
+    let rename_store = ArtifactStore::open(&rename_root.0).unwrap();
+    let stored = rename_store.put(b"rename failure bytes").unwrap();
+    let (_, candidates) = rename_store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
     let candidate = candidates.into_iter().next().unwrap();
-    set_checkpoint("reconciliation_quarantine_source_fsync", FaultAction::Fail);
+    set_checkpoint("reconciliation_quarantine_rename", FaultAction::Fail);
     assert_eq!(
-        store.quarantine_reconciliation(candidate.clone()).unwrap(),
+        rename_store.quarantine_reconciliation(candidate.clone()),
-        ReconciliationMutation::Retryable
+        Err(ArtifactError::Storage)
+    );
+    assert_eq!(
+        rename_store.read(&stored.artifact_ref).unwrap(),
+        b"rename failure bytes"
     );
     clear_checkpoint();
     assert_eq!(
-        store.quarantine_reconciliation(candidate).unwrap(),
+        rename_store.quarantine_reconciliation(candidate).unwrap(),
-        ReconciliationMutation::AlreadyQuarantined
+        ReconciliationMutation::Quarantined
+    );
+
+    let delete_root = TestRoot::new("unlink-fail");
+    let delete_store = ArtifactStore::open(&delete_root.0).unwrap();
+    delete_store.put(b"unlink failure bytes").unwrap();
+    let (_, candidates) = delete_store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    delete_store
+        .quarantine_reconciliation(candidates.into_iter().next().unwrap())
+        .unwrap();
+    let (_, candidates) = delete_store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    let candidate = candidates.into_iter().next().unwrap();
+    set_checkpoint("reconciliation_delete_unlink", FaultAction::Fail);
+    assert_eq!(
+        delete_store.delete_quarantined_reconciliation(candidate.clone()),
+        Err(ArtifactError::Storage)
+    );
+    clear_checkpoint();
+    assert_eq!(
+        delete_store
+            .delete_quarantined_reconciliation(candidate)
+            .unwrap(),
+        ReconciliationMutation::Deleted
     );
 }
 
 #[test]
 #[cfg(debug_assertions)]
-fn storage_fault_during_stat_propagates() {
+fn storage_fault_returns_a_retryable_page_before_the_failed_entry() {
     let _guard = fault_guard();
     let root = TestRoot::new("stat-storage");
     let store = ArtifactStore::open(&root.0).unwrap();
-    store.put(b"reconciliation stat storage").unwrap();
+    let sha = root.0.join("sha256");
-    set_checkpoint("reconciliation_stat", FaultAction::Fail);
+    let shard = sha.join("aa");
-    assert!(matches!(
+    fs::create_dir(&sha).unwrap();
-        store.scan_reconciliation(None, 512, 8),
+    fs::create_dir(&shard).unwrap();
-        Err(ArtifactError::Storage)
+    fs::set_permissions(&sha, fs::Permissions::from_mode(0o700)).unwrap();
-    ));
+    fs::set_permissions(&shard, fs::Permissions::from_mode(0o700)).unwrap();
+    for suffix in ['1', '2'] {
+        let path = shard.join(format!("aa{}", suffix.to_string().repeat(62)));
+        fs::write(&path, b"valid opaque candidate").unwrap();
+        fs::set_permissions(path, fs::Permissions::from_mode(0o400)).unwrap();
+    }
+
+    set_checkpoint_on_hit("reconciliation_stat", FaultAction::Fail, 2);
+    let (report, mut candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::Retryable);
+    assert_eq!(report.scanned, 1);
+    assert_eq!(report.final_entries, 1);
+    assert_eq!(candidates.len(), 1);
+    let continuation = report.continuation;
     clear_checkpoint();
+
+    let (report, resumed) = store
+        .scan_reconciliation(continuation, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::Complete);
+    assert_eq!(report.scanned, 1);
+    assert_eq!(report.final_entries, 1);
+    candidates.extend(resumed);
+    assert_eq!(candidates.len(), 2);
+    for candidate in candidates {
+        assert_eq!(
+            store.quarantine_reconciliation(candidate).unwrap(),
+            ReconciliationMutation::Quarantined
+        );
+    }
+    let (report, quarantined) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.final_entries, 0);
+    assert_eq!(report.quarantined_entries, 2);
+    assert_eq!(quarantined.len(), 2);
+}
+
+#[test]
+#[cfg(debug_assertions)]
+fn getdents_retry_restores_the_pre_call_directory_cookie() {
+    let _guard = fault_guard();
+    let root = TestRoot::new("getdents-storage");
+    let store = ArtifactStore::open(&root.0).unwrap();
+    store.put(b"getdents retry bytes").unwrap();
+
+    set_checkpoint("reconciliation_getdents", FaultAction::Fail);
+    let (report, candidates) = store
+        .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::Retryable);
+    assert!(candidates.is_empty());
+    clear_checkpoint();
+
+    let (report, candidates) = store
+        .scan_reconciliation(report.continuation, 1, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::ScanBudget);
+    assert_eq!(report.traversal_syscalls, 1);
+    assert!(candidates.is_empty());
+
+    let (report, candidates) = store
+        .scan_reconciliation(report.continuation, FULL_SCAN_BUDGET, 8)
+        .unwrap();
+    assert_eq!(report.stop, ReconciliationScanStop::Complete);
+    assert_eq!(candidates.len(), 1);
 }
 
 #[test]
@@ -370,7 +703,9 @@ fn crash_windows_are_recoverable() {
             FaultAction::Exit,
         );
         let store = ArtifactStore::open(PathBuf::from(root)).unwrap();
-        let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+        let (_, candidates) = store
+            .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+            .unwrap();
         let candidate = candidates.into_iter().next().unwrap();
         if action == "quarantine" {
             let _ = store.quarantine_reconciliation(candidate);
@@ -391,7 +726,9 @@ fn crash_windows_are_recoverable() {
         let store = ArtifactStore::open(&root.0).unwrap();
         let stored = store.put(b"reconciliation crash bytes").unwrap();
         if action == "delete" {
-            let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+            let (_, candidates) = store
+                .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+                .unwrap();
             store
                 .quarantine_reconciliation(candidates.into_iter().next().unwrap())
                 .unwrap();
@@ -406,7 +743,9 @@ fn crash_windows_are_recoverable() {
         assert_eq!(status.code(), Some(86), "stage={stage}");
 
         let store = ArtifactStore::open(&root.0).unwrap();
-        let (_, candidates) = store.scan_reconciliation(None, 512, 8).unwrap();
+        let (_, candidates) = store
+            .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+            .unwrap();
         for candidate in candidates {
             match candidate.namespace() {
                 ReconciliationNamespace::Final => {
@@ -417,7 +756,9 @@ fn crash_windows_are_recoverable() {
                 }
             }
         }
-        let (_, remaining) = store.scan_reconciliation(None, 512, 8).unwrap();
+        let (_, remaining) = store
+            .scan_reconciliation(None, FULL_SCAN_BUDGET, 8)
+            .unwrap();
         assert!(remaining.is_empty(), "stage={stage}");
         assert_eq!(
             store