test(kms): validate Vault circuit recovery

* fix(scanner): isolate corrupt cycle state

* fix(scanner): preserve newer state during recovery reset

* fix(scanner): fence recovery reset state

* fix(scanner): reject terminal recovery epochs

* fix(scanner): reject trailing cycle state bytes

* fix(scanner): reject terminal leadership epochs

* fix(scanner): retain recovery wake notifications

* fix(scanner): recover from oversized markers

.github/workflows/nightly-gnu.yml

@@ -39,11 +39,10 @@ jobs:
     env:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
     steps:
-      - name: Checkout main branch
+      - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
-          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup
@@ -89,11 +88,10 @@ jobs:
       # either casing.
       NO_PROXY: 127.0.0.1,localhost
     steps:
-      - name: Checkout main branch
+      - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
-          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup
@@ -178,11 +176,10 @@ jobs:
       FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
       NO_PROXY: 127.0.0.1,localhost
     steps:
-      - name: Checkout main branch
+      - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
-          ref: main
 
       - name: Setup Rust environment
         uses: ./.github/actions/setup

crates/filemeta/src/metacache.rs

@@ -21,7 +21,6 @@ use arc_swap::ArcSwapOption;
 use rmp::Marker;
 use serde::{Deserialize, Serialize};
 use std::cmp::Ordering;
-use std::collections::{HashMap, HashSet};
 use std::str::from_utf8;
 use std::{
     fmt::Debug,
@@ -38,13 +37,8 @@ use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
 use tokio::spawn;
 use tokio::sync::Mutex;
 use tracing::{debug, warn};
-use uuid::Uuid;
 
 const SLASH_SEPARATOR: &str = "/";
-pub const MAX_META_CACHE_HEAL_CANDIDATES: usize = 1024;
-/// Keep truncation continuations bounded while still giving the scanner a
-/// safe object-level retry for versions that did not fit in the candidate set.
-pub const MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS: usize = 64;
 
 #[derive(Clone, Debug, Default)]
 pub struct MetadataResolutionParams {
@@ -72,47 +66,6 @@ pub struct MetaCacheEntry {
     pub reusable: bool,
 }
 
-#[derive(Clone, Debug, PartialEq, Eq, Hash)]
-pub enum MetaCacheHealCandidateKind {
-    Object,
-    DeleteMarker,
-    UnversionedObject,
-}
-
-#[derive(Clone, Debug, PartialEq, Eq, Hash)]
-pub struct MetaCacheHealCandidate {
-    pub object: String,
-    pub version_id: Option<Uuid>,
-    pub kind: MetaCacheHealCandidateKind,
-    /// Number of raw disk entries that carried this validated version.
-    pub replica_count: usize,
-}
-
-impl MetaCacheHealCandidate {
-    pub fn validated_version(&self) -> Option<Uuid> {
-        match self.kind {
-            MetaCacheHealCandidateKind::Object | MetaCacheHealCandidateKind::DeleteMarker => self.version_id,
-            MetaCacheHealCandidateKind::UnversionedObject => None,
-        }
-    }
-
-    pub fn is_unversioned(&self) -> bool {
-        self.kind == MetaCacheHealCandidateKind::UnversionedObject
-    }
-}
-
-#[derive(Clone, Debug, Default, PartialEq, Eq)]
-pub struct MetaCacheHealDiscovery {
-    pub candidates: Vec<MetaCacheHealCandidate>,
-    pub unverified_count: usize,
-    pub truncated: bool,
-    /// Object names whose validated version set exceeded the candidate cap.
-    /// The scanner retries these names without a version and with destructive
-    /// healing disabled; this is an explicit bounded continuation, not a
-    /// version claim.
-    pub truncated_objects: Vec<String>,
-}
-
 impl MetaCacheEntry {
     pub fn marshal_msg(&self) -> Result<Vec<u8>> {
         let mut wr = Vec::new();
@@ -417,184 +370,6 @@ impl MetaCacheEntries {
         })
     }
 
-    /// Discover validated object/delete-marker versions and safe unversioned
-    /// inspection candidates in the raw entries without applying read quorum.
-    /// This is intentionally separate from [`Self::resolve`]: a sub-quorum
-    /// version is a valid heal target even though it must not participate in
-    /// normal reads or writes.
-    ///
-    /// The validated list is bounded and deduplicated by object, version id,
-    /// and metadata kind; each candidate retains the number of raw disk
-    /// entries that carried it so callers can classify sub-quorum versions.
-    /// Entries whose xl.meta cannot be decoded are counted separately for
-    /// discovery accounting; they never become versionless destructive heal
-    /// requests and do not consume the validated quota. An
-    /// [`MetaCacheHealCandidateKind::UnversionedObject`] is always consumed by
-    /// a non-destructive scanner request.
-    pub fn discover_heal_candidates(&self, bucket: &str, max_candidates: usize) -> MetaCacheHealDiscovery {
-        let limit = max_candidates.min(MAX_META_CACHE_HEAL_CANDIDATES);
-        if limit == 0 || bucket.is_empty() {
-            return MetaCacheHealDiscovery::default();
-        }
-
-        let mut discovery = MetaCacheHealDiscovery {
-            candidates: Vec::<MetaCacheHealCandidate>::with_capacity(limit.min(self.0.len())),
-            unverified_count: 0,
-            truncated: false,
-            truncated_objects: Vec::with_capacity(MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS.min(limit)),
-        };
-        let mut seen: HashMap<(String, Option<Uuid>, MetaCacheHealCandidateKind), usize> =
-            HashMap::with_capacity(limit.min(self.0.len()));
-
-        for entry in self.0.iter().flatten() {
-            if !valid_heal_candidate_name(bucket, entry) {
-                continue;
-            }
-
-            let meta = match FileMeta::load(&entry.metadata) {
-                Ok(meta) => meta,
-                Err(_) => {
-                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                    continue;
-                }
-            };
-            let mut entry_seen = HashSet::new();
-
-            for shallow in meta.versions {
-                let version = match shallow.parse_version_meta() {
-                    Ok(version) if version.valid() => version,
-                    Ok(_) | Err(_) => {
-                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                        continue;
-                    }
-                };
-                if version.free_version() {
-                    continue;
-                }
-
-                let payload_header = version.header();
-                if normalize_version_id(shallow.header.version_id) != normalize_version_id(payload_header.version_id)
-                    || shallow.header.version_type != payload_header.version_type
-                {
-                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                    continue;
-                }
-
-                let (kind, version_id) = match version.version_type {
-                    VersionType::Object
-                        if version.object.is_some() && version.delete_marker.is_none() && version.legacy_object.is_none() =>
-                    {
-                        match version.object.as_ref().and_then(|object| object.version_id) {
-                            Some(id) if !id.is_nil() => (MetaCacheHealCandidateKind::Object, Some(id)),
-                            Some(_) | None => (MetaCacheHealCandidateKind::UnversionedObject, None),
-                        }
-                    }
-                    VersionType::Delete
-                        if version.delete_marker.is_some() && version.object.is_none() && version.legacy_object.is_none() =>
-                    {
-                        let Some(id) = version.delete_marker.as_ref().and_then(|marker| marker.version_id) else {
-                            discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                            continue;
-                        };
-                        if id.is_nil() {
-                            discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                            continue;
-                        }
-                        (MetaCacheHealCandidateKind::DeleteMarker, Some(id))
-                    }
-                    VersionType::Legacy
-                        if version.legacy_object.is_some() && version.object.is_none() && version.delete_marker.is_none() =>
-                    {
-                        let Some(legacy) = version.legacy_object.as_ref() else {
-                            continue;
-                        };
-                        if legacy.version_id.is_empty() {
-                            (MetaCacheHealCandidateKind::UnversionedObject, None)
-                        } else {
-                            let Ok(id) = Uuid::parse_str(&legacy.version_id) else {
-                                discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                                continue;
-                            };
-                            if id.is_nil() {
-                                discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                                continue;
-                            }
-                            (MetaCacheHealCandidateKind::Object, Some(id))
-                        }
-                    }
-                    _ => {
-                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                        continue;
-                    }
-                };
-
-                if normalize_version_id(payload_header.version_id) != version_id {
-                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                    continue;
-                }
-
-                // `all_parts=true` is the trust-boundary check for versioned
-                // candidates. A null/legacy object may still need the old
-                // non-destructive inspection fallback when its part arrays
-                // are parseable but incomplete; never use that fallback for
-                // a candidate carrying a real version id.
-                let file_info = match version.clone().into_fileinfo(bucket, &entry.name, true) {
-                    Ok(file_info) => file_info,
-                    Err(_) if version_id.is_none() && matches!(kind, MetaCacheHealCandidateKind::UnversionedObject) => {
-                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                        match version.into_fileinfo(bucket, &entry.name, false) {
-                            Ok(file_info) => file_info,
-                            Err(_) => continue,
-                        }
-                    }
-                    Err(_) => {
-                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                        continue;
-                    }
-                };
-                if file_info.volume != bucket || file_info.name != entry.name {
-                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
-                    continue;
-                }
-
-                let candidate = MetaCacheHealCandidate {
-                    object: entry.name.clone(),
-                    version_id,
-                    kind,
-                    replica_count: 1,
-                };
-                let key = (candidate.object.clone(), candidate.version_id, candidate.kind.clone());
-                if entry_seen.contains(&key) {
-                    continue;
-                }
-                if let Some(index) = seen.get(&key).copied() {
-                    entry_seen.insert(key);
-                    discovery.candidates[index].replica_count = discovery.candidates[index].replica_count.saturating_add(1);
-                } else if discovery.candidates.len() >= limit {
-                    // Keep the validated candidate list bounded, but retain a
-                    // bounded object-level continuation so the scanner cannot
-                    // silently lose every version of a busy object.
-                    discovery.truncated = true;
-                    if discovery.truncated_objects.len() < MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
-                        && !discovery.truncated_objects.iter().any(|object| object == &candidate.object)
-                    {
-                        discovery.truncated_objects.push(candidate.object.clone());
-                    }
-                    // The remaining versions in this raw entry cannot add a
-                    // bounded candidate; avoid parsing a very long history
-                    // after the safe continuation has been recorded.
-                    break;
-                } else {
-                    entry_seen.insert(key.clone());
-                    seen.insert(key, discovery.candidates.len());
-                    discovery.candidates.push(candidate);
-                }
-            }
-        }
-
-        discovery
-    }
-
     fn resolve_inner(&self, mut params: MetadataResolutionParams, enforce_write_quorum: bool) -> Option<MetaCacheEntry> {
         if self.0.is_empty() {
             debug!(
@@ -771,33 +546,6 @@ impl MetaCacheEntries {
     }
 }
 
-fn valid_heal_candidate_name(bucket: &str, entry: &MetaCacheEntry) -> bool {
-    if bucket.is_empty()
-        || entry.name.is_empty()
-        || entry.is_dir()
-        || entry.name.contains('\\')
-        || entry.name.chars().any(char::is_control)
-    {
-        return false;
-    }
-
-    // Validate raw key components without normalizing them. The scanner maps
-    // accepted keys to filesystem paths later, so dot components and empty
-    // internal components must be rejected before that boundary. A final
-    // empty component is retained for valid keys ending in '/'.
-    let mut components = entry.name.split('/').peekable();
-    while let Some(component) = components.next() {
-        if component == "." || component == ".." || (component.is_empty() && components.peek().is_some()) {
-            return false;
-        }
-    }
-    true
-}
-
-fn normalize_version_id(version_id: Option<Uuid>) -> Option<Uuid> {
-    version_id.filter(|id| !id.is_nil())
-}
-
 #[derive(Debug, Default)]
 pub struct MetaCacheEntriesSortedResult {
     pub entries: Option<MetaCacheEntriesSorted>,
@@ -1243,7 +991,7 @@ impl<T: Clone + Debug + Send + Sync + 'static> Cache<T> {
 mod tests {
     use super::*;
     use crate::test_data::create_real_xlmeta;
-    use crate::{FileMetaVersion, MetaDeleteMarker, MetaObjectV1, MetaObjectV1Erasure, MetaObjectV1Stat, TRANSITION_COMPLETE};
+    use crate::{FileMetaVersion, MetaDeleteMarker, TRANSITION_COMPLETE};
     use std::collections::HashMap;
     use std::io::Cursor;
     use std::sync::{
@@ -1844,348 +1592,6 @@ mod tests {
         );
     }
 
-    #[test]
-    fn discover_heal_candidates_keeps_sub_quorum_versions_and_deduplicates() {
-        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
-        let entries = MetaCacheEntries(vec![
-            Some(metacache_entry_single_version(1, now, "one")),
-            Some(metacache_entry_single_version(2, now, "two")),
-            Some(metacache_entry_single_version(2, now, "two")),
-            Some(metacache_entry_single_version(3, now, "three")),
-        ]);
-
-        let discovery = entries.discover_heal_candidates("bucket", 16);
-        let ids: std::collections::HashSet<Uuid> = discovery
-            .candidates
-            .iter()
-            .filter_map(|candidate| candidate.version_id)
-            .collect();
-        assert_eq!(
-            ids,
-            [Uuid::from_u128(1), Uuid::from_u128(2), Uuid::from_u128(3)]
-                .into_iter()
-                .collect()
-        );
-        assert_eq!(discovery.candidates.len(), 3, "duplicate tied versions must be emitted once");
-        assert_eq!(
-            discovery
-                .candidates
-                .iter()
-                .find(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
-                .expect("duplicate version should be discovered")
-                .replica_count,
-            2
-        );
-    }
-
-    #[test]
-    fn discover_heal_candidates_does_not_count_duplicate_versions_within_one_entry() {
-        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
-        let mut meta = FileMeta::load(&metacache_entry_single_version(1, now, "duplicate").metadata)
-            .expect("duplicate fixture should decode");
-        meta.versions.push(meta.versions[0].clone());
-        let entry = MetaCacheEntry {
-            name: "object".to_string(),
-            metadata: meta.marshal_msg().expect("duplicate metadata should marshal"),
-            cached: Some(meta),
-            reusable: false,
-        };
-
-        let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 16);
-        let candidate = discovery
-            .candidates
-            .iter()
-            .find(|candidate| candidate.version_id == Some(Uuid::from_u128(1)))
-            .expect("duplicate fixture should be discovered");
-        assert_eq!(candidate.replica_count, 1);
-    }
-
-    #[test]
-    fn discover_heal_candidates_covers_divergent_quorum_boundaries_n2_n4_n6() {
-        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
-
-        for (disk_count, quorum) in [(2usize, 1usize), (4, 2), (6, 3)] {
-            let target_id = Uuid::from_u128(0x1000 + disk_count as u128);
-            for target_replicas in [quorum.saturating_sub(1), quorum, quorum + 1] {
-                let entries = (0..disk_count)
-                    .map(|disk| {
-                        let version_id = if disk < target_replicas {
-                            target_id
-                        } else {
-                            Uuid::from_u128(0x2000 + disk as u128)
-                        };
-                        Some(metacache_entry_single_version(version_id.as_u128(), now, "divergent"))
-                    })
-                    .collect();
-                let discovery = MetaCacheEntries(entries).discover_heal_candidates("bucket", 32);
-                let target = discovery
-                    .candidates
-                    .iter()
-                    .find(|candidate| candidate.version_id == Some(target_id));
-                assert_eq!(target.is_some(), target_replicas > 0, "N={disk_count}, replicas={target_replicas}");
-                if let Some(target) = target {
-                    assert_eq!(target.replica_count, target_replicas);
-                }
-            }
-        }
-    }
-
-    #[test]
-    fn discover_heal_candidates_separates_delete_markers_and_preserves_unversioned_objects() {
-        let mut marker_meta = FileMeta::new();
-        marker_meta
-            .add_version(FileInfo {
-                volume: "bucket".to_string(),
-                name: "object".to_string(),
-                version_id: Some(Uuid::from_u128(99)),
-                deleted: true,
-                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
-                ..Default::default()
-            })
-            .expect("delete marker should be added");
-        let marker = MetaCacheEntry {
-            name: "object".to_string(),
-            metadata: marker_meta.marshal_msg().expect("delete marker metadata should marshal"),
-            cached: Some(marker_meta),
-            reusable: false,
-        };
-
-        let unversioned_entry = metacache_entry_with_mod_time(
-            OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp"),
-            "unversioned",
-        );
-        let discovery = MetaCacheEntries(vec![Some(marker), Some(unversioned_entry)]).discover_heal_candidates("bucket", 16);
-        assert!(discovery.candidates.iter().any(|candidate| {
-            candidate.kind == MetaCacheHealCandidateKind::DeleteMarker && candidate.version_id == Some(Uuid::from_u128(99))
-        }));
-        assert!(discovery.candidates.iter().any(|candidate| {
-            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
-        }));
-    }
-
-    #[test]
-    fn discover_heal_candidates_rejects_delete_markers_without_ids() {
-        let mut marker_meta = FileMeta::new();
-        marker_meta
-            .add_version(FileInfo {
-                volume: "bucket".to_string(),
-                name: "object".to_string(),
-                deleted: true,
-                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
-                ..Default::default()
-            })
-            .expect("nil delete marker should be added");
-        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
-            name: "object".to_string(),
-            metadata: marker_meta.marshal_msg().expect("nil marker metadata should marshal"),
-            cached: Some(marker_meta),
-            reusable: false,
-        })])
-        .discover_heal_candidates("bucket", 16);
-
-        assert!(
-            !discovery
-                .candidates
-                .iter()
-                .any(|candidate| candidate.kind == MetaCacheHealCandidateKind::DeleteMarker)
-        );
-        assert!(discovery.unverified_count >= 1);
-    }
-
-    #[test]
-    fn discover_heal_candidates_skips_free_versions() {
-        let object_id = Uuid::from_u128(100);
-        let free_id = Uuid::from_u128(101);
-        let mut meta = FileMeta::new();
-        meta.add_version(FileInfo {
-            volume: "bucket".to_string(),
-            name: "object".to_string(),
-            version_id: Some(object_id),
-            transition_status: TRANSITION_COMPLETE.to_string(),
-            transitioned_objname: "remote/object".to_string(),
-            transition_version_id: Some(Uuid::from_u128(102)),
-            transition_tier: "WARM".to_string(),
-            mod_time: Some(OffsetDateTime::now_utc()),
-            ..Default::default()
-        })
-        .expect("transitioned object should be added");
-        let mut delete = FileInfo {
-            volume: "bucket".to_string(),
-            name: "object".to_string(),
-            version_id: Some(object_id),
-            mod_time: Some(OffsetDateTime::now_utc()),
-            ..Default::default()
-        };
-        delete.set_tier_free_version_id(&free_id.to_string());
-        meta.delete_version(&delete).expect("free version should be persisted");
-
-        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
-            name: "object".to_string(),
-            metadata: meta.marshal_msg().expect("free version metadata should marshal"),
-            cached: Some(meta),
-            reusable: false,
-        })])
-        .discover_heal_candidates("bucket", 16);
-        assert!(discovery.candidates.is_empty());
-    }
-
-    #[test]
-    fn discover_heal_candidates_preserves_unversioned_legacy_object() {
-        let legacy = MetaObjectV1 {
-            version: "1.0.1".to_string(),
-            format: "xl".to_string(),
-            stat: MetaObjectV1Stat {
-                size: 1,
-                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
-                name: "object".to_string(),
-                ..Default::default()
-            },
-            erasure: MetaObjectV1Erasure {
-                data_blocks: 4,
-                parity_blocks: 2,
-                index: 1,
-                distribution: vec![1, 2, 3, 4, 5, 6],
-                ..Default::default()
-            },
-            ..Default::default()
-        };
-        let version = FileMetaVersion {
-            version_type: VersionType::Legacy,
-            legacy_object: Some(legacy),
-            ..Default::default()
-        };
-        let mut meta = FileMeta::new();
-        meta.versions
-            .push(FileMetaShallowVersion::try_from(version).expect("legacy metadata should marshal"));
-        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
-            name: "object".to_string(),
-            metadata: meta.marshal_msg().expect("legacy metadata should marshal"),
-            cached: Some(meta),
-            reusable: false,
-        })])
-        .discover_heal_candidates("bucket", 16);
-        assert!(discovery.candidates.iter().any(|candidate| {
-            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
-        }));
-    }
-
-    #[test]
-    fn discover_heal_candidates_rejects_nil_and_malformed_metadata_and_is_bounded() {
-        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
-        let mut nil = metacache_entry_single_version(1, now, "nil");
-        let mut nil_meta = FileMeta::load(&nil.metadata).expect("nil fixture should decode");
-        let mut nil_version = nil_meta.versions[0]
-            .parse_version_meta()
-            .expect("nil fixture version should decode");
-        nil_version.object.as_mut().expect("object fixture").version_id = Some(Uuid::nil());
-        nil_meta.versions[0] = FileMetaShallowVersion::try_from(nil_version).expect("nil fixture should marshal");
-        nil.metadata = nil_meta.marshal_msg().expect("nil fixture metadata should marshal");
-
-        let mut mismatched = metacache_entry_single_version(2, now, "mismatched");
-        let mut mismatched_meta = FileMeta::load(&mismatched.metadata).expect("mismatched fixture should decode");
-        mismatched_meta.versions[0].header.version_id = Some(Uuid::from_u128(200));
-        mismatched.metadata = mismatched_meta.marshal_msg().expect("mismatched metadata should marshal");
-
-        let mut short_parts = metacache_entry_single_version(3, now, "short-parts");
-        let mut short_parts_meta = FileMeta::load(&short_parts.metadata).expect("short-parts fixture should decode");
-        let mut short_parts_version = short_parts_meta.versions[0]
-            .parse_version_meta()
-            .expect("short-parts fixture version should decode");
-        let object = short_parts_version.object.as_mut().expect("object fixture");
-        object.part_numbers = vec![1];
-        object.part_actual_sizes = vec![1];
-        object.part_sizes.clear();
-        short_parts_meta.versions[0] = FileMetaShallowVersion::try_from(short_parts_version).expect("short-parts should marshal");
-        short_parts.metadata = short_parts_meta.marshal_msg().expect("short-parts metadata should marshal");
-
-        let mut short_unversioned = metacache_entry_with_mod_time(now, "short-unversioned");
-        let mut short_unversioned_meta =
-            FileMeta::load(&short_unversioned.metadata).expect("short-unversioned fixture should decode");
-        let mut short_unversioned_version = short_unversioned_meta.versions[0]
-            .parse_version_meta()
-            .expect("short-unversioned version should decode");
-        let unversioned_object = short_unversioned_version.object.as_mut().expect("unversioned object fixture");
-        unversioned_object.part_numbers = vec![1];
-        unversioned_object.part_actual_sizes = vec![1];
-        unversioned_object.part_sizes.clear();
-        short_unversioned_meta.versions[0] =
-            FileMetaShallowVersion::try_from(short_unversioned_version).expect("short-unversioned should marshal");
-        short_unversioned.metadata = short_unversioned_meta
-            .marshal_msg()
-            .expect("short-unversioned metadata should marshal");
-
-        let mut malformed = nil.clone();
-        malformed.name = "malformed".to_string();
-        malformed.metadata = vec![1, 2, 3];
-
-        let entries = MetaCacheEntries(
-            std::iter::once(Some(nil))
-                .chain(std::iter::once(Some(mismatched)))
-                .chain(std::iter::once(Some(short_parts)))
-                .chain(std::iter::once(Some(short_unversioned)))
-                .chain(std::iter::once(Some(malformed)))
-                .chain((0..32).map(|id| Some(metacache_entry_single_version(id + 10, now, "bounded"))))
-                .collect(),
-        );
-        let discovery = entries.discover_heal_candidates("bucket", 5);
-        assert!(discovery.candidates.len() <= 5);
-        assert!(discovery.truncated, "bounded discovery must expose dropped candidates");
-        assert!(
-            discovery.truncated_objects.iter().any(|object| object == "object"),
-            "bounded discovery must expose an object-level safe continuation"
-        );
-        assert!(
-            !discovery
-                .candidates
-                .iter()
-                .any(|candidate| candidate.version_id == Some(Uuid::nil()))
-        );
-        assert!(
-            !discovery
-                .candidates
-                .iter()
-                .any(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
-        );
-        assert!(
-            !discovery
-                .candidates
-                .iter()
-                .any(|candidate| candidate.version_id == Some(Uuid::from_u128(3)))
-        );
-        assert!(discovery.candidates.iter().any(|candidate| {
-            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
-        }));
-        assert!(
-            discovery.unverified_count >= 1,
-            "malformed and rejected metadata must remain observable during discovery"
-        );
-
-        for invalid_name in [
-            "../object",
-            "./object",
-            "object/../other",
-            "object//name",
-            "object\\name",
-            "object\u{0001}name",
-            "object\0name",
-        ] {
-            let mut entry = metacache_entry_single_version(400, now, invalid_name);
-            entry.name = invalid_name.to_string();
-            let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
-            assert!(
-                discovery.candidates.is_empty(),
-                "invalid key should not become a heal candidate: {invalid_name:?}"
-            );
-        }
-
-        for valid_name in ["trailing/", "prefix/object"] {
-            let mut entry = metacache_entry_single_version(401, now, valid_name);
-            entry.name = valid_name.to_string();
-            let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
-            assert_eq!(discovery.candidates.len(), 1, "raw S3 key should remain opaque: {valid_name:?}");
-        }
-    }
-
     #[test]
     fn resolve_rejects_partial_latest_and_returns_committed_previous_metadata() {
         let old_mod_time = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");

crates/kms/tests/vault_ha_failover_live.rs

@@ -16,14 +16,14 @@
 //!
 //! `scripts/test/vault_ha_kms_live.sh` owns the official Vault containers and
 //! kills the active node while this test continuously decrypts through a
-//! surviving standby. KV2 and Transit requests must remain successful, use a
-//! bounded number of attempts, and leave the circuit and in-flight gauges at
-//! zero after a new leader is elected.
+//! surviving standby. KV2 and Transit must recover after the bounded circuit
+//! interval, use a bounded number of attempts, and leave the circuit and
+//! in-flight gauges at zero after a new leader is elected.
 
 use std::collections::HashMap;
 use std::path::{Path, PathBuf};
-use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
 use metrics_util::MetricKind;
@@ -43,6 +43,11 @@ const OPERATION_ATTEMPTS: &str = "rustfs_kms_backend_operation_attempts";
 const IN_FLIGHT: &str = "rustfs_kms_backend_in_flight";
 const CIRCUIT_OPEN: &str = "rustfs_kms_backend_circuit_open";
 const MAX_ATTEMPTS: u32 = 10;
+const ATTEMPT_TIMEOUT: Duration = Duration::from_secs(2);
+const HEALTHY_PROGRESS_TIMEOUT: Duration = Duration::from_secs(20);
+// The circuit remains open for 30s after five failed attempts.
+const POST_FAILOVER_PROGRESS_TIMEOUT: Duration = Duration::from_secs(35);
+const FAILOVER_ERROR_POLL_INTERVAL: Duration = Duration::from_millis(100);
 
 type MetricEntry = (
     metrics_util::CompositeKey,
@@ -64,7 +69,7 @@ fn config(backend: KmsBackend, backend_config: BackendConfig) -> KmsConfig {
         backend,
         backend_config,
         allow_insecure_dev_defaults: true,
-        timeout: Duration::from_secs(2),
+        timeout: ATTEMPT_TIMEOUT,
         retry_attempts: MAX_ATTEMPTS,
         enable_cache: false,
         ..KmsConfig::default()
@@ -164,14 +169,31 @@ fn retryable_failures(snapshot: &[MetricEntry], operation: &str) -> u64 {
         .sum()
 }
 
-async fn wait_for_count(counter: &AtomicU64, minimum: u64, description: &str) {
-    tokio::time::timeout(Duration::from_secs(20), async {
+async fn wait_for_count(
+    counter: &AtomicU64,
+    failure: &Mutex<Option<String>>,
+    minimum: u64,
+    description: &str,
+    timeout: Duration,
+) {
+    tokio::time::timeout(timeout, async {
         while counter.load(Ordering::SeqCst) < minimum {
+            if let Some(error) = failure.lock().expect("decrypt failure lock poisoned").as_ref() {
+                panic!(
+                    "{description} worker failed after {} successful decrypts: {error}",
+                    counter.load(Ordering::SeqCst)
+                );
+            }
             tokio::time::sleep(Duration::from_millis(25)).await;
         }
     })
     .await
-    .unwrap_or_else(|_| panic!("timed out waiting for {description}"));
+    .unwrap_or_else(|_| {
+        panic!(
+            "timed out after {timeout:?} waiting for {description}: completed {}, expected {minimum}",
+            counter.load(Ordering::SeqCst)
+        )
+    });
 }
 
 async fn wait_for_file(path: &Path, description: &str) {
@@ -189,7 +211,8 @@ async fn decrypt_loop<B: KmsBackendTrait + Send + Sync + 'static>(
     request: DecryptRequest,
     expected: Vec<u8>,
     completed: Arc<AtomicU64>,
-    failed: Arc<AtomicBool>,
+    allow_failover_errors: Arc<AtomicBool>,
+    failure: Arc<Mutex<Option<String>>>,
     stop: CancellationToken,
 ) {
     while !stop.is_cancelled() {
@@ -197,8 +220,18 @@ async fn decrypt_loop<B: KmsBackendTrait + Send + Sync + 'static>(
             Ok(response) if response.plaintext == expected => {
                 completed.fetch_add(1, Ordering::SeqCst);
             }
-            Ok(_) | Err(_) => {
-                failed.store(true, Ordering::SeqCst);
+            Ok(_) => {
+                *failure.lock().expect("decrypt failure lock poisoned") =
+                    Some("decrypt returned unexpected plaintext".to_string());
+                return;
+            }
+            Err(rustfs_kms::KmsError::BackendError { .. } | rustfs_kms::KmsError::OperationTimedOut { .. })
+                if allow_failover_errors.load(Ordering::SeqCst) =>
+            {
+                tokio::time::sleep(FAILOVER_ERROR_POLL_INTERVAL).await;
+            }
+            Err(error) => {
+                *failure.lock().expect("decrypt failure lock poisoned") = Some(error.to_string());
                 return;
             }
         }
@@ -296,7 +329,9 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
     );
 
     let stop = CancellationToken::new();
-    let failed = Arc::new(AtomicBool::new(false));
+    let allow_failover_errors = Arc::new(AtomicBool::new(false));
+    let kv2_failure = Arc::new(Mutex::new(None));
+    let transit_failure = Arc::new(Mutex::new(None));
     let kv2_completed = Arc::new(AtomicU64::new(0));
     let transit_completed = Arc::new(AtomicU64::new(0));
     let kv2_worker = tokio::spawn(decrypt_loop(
@@ -304,7 +339,8 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
         kv2_request,
         kv2_data_key.plaintext_key,
         Arc::clone(&kv2_completed),
-        Arc::clone(&failed),
+        Arc::clone(&allow_failover_errors),
+        Arc::clone(&kv2_failure),
         stop.clone(),
     ));
     let transit_worker = tokio::spawn(decrypt_loop(
@@ -312,12 +348,21 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
         transit_request,
         transit_data_key.plaintext_key,
         Arc::clone(&transit_completed),
-        Arc::clone(&failed),
+        Arc::clone(&allow_failover_errors),
+        Arc::clone(&transit_failure),
         stop.clone(),
     ));
 
-    wait_for_count(&kv2_completed, 2, "two healthy KV2 decrypts").await;
-    wait_for_count(&transit_completed, 2, "two healthy Transit decrypts").await;
+    wait_for_count(&kv2_completed, &kv2_failure, 2, "two healthy KV2 decrypts", HEALTHY_PROGRESS_TIMEOUT).await;
+    wait_for_count(
+        &transit_completed,
+        &transit_failure,
+        2,
+        "two healthy Transit decrypts",
+        HEALTHY_PROGRESS_TIMEOUT,
+    )
+    .await;
+    allow_failover_errors.store(true, Ordering::SeqCst);
     std::fs::write(&marker, b"ready").expect("publish failover readiness marker");
 
     wait_for_file(&elected, "the replacement Vault leader").await;
@@ -326,18 +371,39 @@ async fn exercise_failover(snapshotter: &Snapshotter) {
 
     let kv2_after_election = kv2_completed.load(Ordering::SeqCst) + 2;
     let transit_after_election = transit_completed.load(Ordering::SeqCst) + 2;
-    wait_for_count(&kv2_completed, kv2_after_election, "post-failover KV2 decrypts").await;
-    wait_for_count(&transit_completed, transit_after_election, "post-failover Transit decrypts").await;
+    wait_for_count(
+        &kv2_completed,
+        &kv2_failure,
+        kv2_after_election,
+        "post-failover KV2 decrypts",
+        POST_FAILOVER_PROGRESS_TIMEOUT,
+    )
+    .await;
+    wait_for_count(
+        &transit_completed,
+        &transit_failure,
+        transit_after_election,
+        "post-failover Transit decrypts",
+        POST_FAILOVER_PROGRESS_TIMEOUT,
+    )
+    .await;
 
     stop.cancel();
     kv2_worker.await.expect("KV2 decrypt worker must join");
     transit_worker.await.expect("Transit decrypt worker must join");
-    assert!(!failed.load(Ordering::SeqCst), "no decrypt may fail or return different plaintext");
+    assert!(
+        kv2_failure.lock().expect("KV2 failure lock poisoned").is_none(),
+        "no KV2 decrypt may fail or return different plaintext"
+    );
+    assert!(
+        transit_failure.lock().expect("Transit failure lock poisoned").is_none(),
+        "no Transit decrypt may fail or return different plaintext"
+    );
 }
 
 #[test]
 #[ignore = "requires a real three-node Vault Raft cluster; run scripts/test/vault_ha_kms_live.sh"]
-fn vault_raft_leader_failure_preserves_kv2_and_transit_decrypts() {
+fn vault_raft_leader_failure_recovers_kv2_and_transit_decrypts() {
     let recorder = DebuggingRecorder::new();
     let snapshotter = recorder.snapshotter();
     metrics::with_local_recorder(&recorder, || {
@@ -349,11 +415,6 @@ fn vault_raft_leader_failure_preserves_kv2_and_transit_decrypts() {
     });
     let snapshot = snapshotter.snapshot().into_vec();
 
-    assert_eq!(
-        counter_value(&snapshot, OPERATIONS_TOTAL, &[("outcome", "circuit_open")]),
-        0,
-        "a bounded leader election must not open the circuit"
-    );
     assert_eq!(
         counter_value(&snapshot, OPERATIONS_TOTAL, &[("outcome", "budget_exhausted")]),
         0,

crates/scanner/src/data_usage_define.rs

@@ -125,6 +125,34 @@ pub(crate) async fn read_config_with_revision<S: ScannerObjectIO>(
     }
 }
 
+/// Read only the object revision without materializing its body.
+pub(crate) async fn read_config_revision<S: ScannerObjectIO>(store: Arc<S>, path: &str) -> StorageResult<DataUsageCacheRevision> {
+    match store
+        .get_object_reader(
+            RUSTFS_META_BUCKET,
+            path,
+            None,
+            HeaderMap::new(),
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+    {
+        Ok(reader) => reader
+            .object_info
+            .etag
+            .filter(|etag| !etag.is_empty())
+            .map(DataUsageCacheRevision::Etag)
+            .ok_or_else(|| StorageError::other(format!("scanner config object {path} has no ETag"))),
+        Err(Error::FileNotFound | Error::VolumeNotFound | Error::ObjectNotFound(_, _) | Error::BucketNotFound(_)) => {
+            Ok(DataUsageCacheRevision::Missing)
+        }
+        Err(err) => Err(err),
+    }
+}
+
 #[derive(Clone, Debug)]
 pub(crate) struct DataUsageCacheRevisions {
     main: DataUsageCacheRevision,
@@ -146,6 +174,11 @@ pub static LEGACY_DATA_USAGE_OBJ_NAME_PATH: LazyLock<String> =
 pub static DATA_USAGE_BLOOM_NAME_PATH: LazyLock<String> =
     LazyLock::new(|| format!("{BUCKET_META_PREFIX}{SLASH_SEPARATOR}{DATA_USAGE_BLOOM_NAME}"));
 
+/// Durable companion object for a cycle-state object which cannot be decoded.
+/// The primary object is deliberately never replaced or deleted by recovery.
+pub static DATA_USAGE_BLOOM_RECOVERY_PATH: LazyLock<String> =
+    LazyLock::new(|| format!("{}.recovery-required.json", DATA_USAGE_BLOOM_NAME_PATH.as_str()));
+
 pub static BACKGROUND_HEAL_INFO_PATH: LazyLock<String> =
     LazyLock::new(|| format!("{BUCKET_META_PREFIX}{SLASH_SEPARATOR}.background-heal.json"));
 

crates/scanner/src/data_usage_define/persistence.rs

@@ -74,7 +74,7 @@ impl DataUsageCache {
         let loaded = Self::load_cache(store.clone(), name).await?;
         let backup = match loaded.backup_revision {
             Some(revision) => Some(revision),
-            None => match Self::revision_for_path(store, &backup_path).await {
+            None => match read_config_revision(store, &backup_path).await {
                 Ok(revision) => Some(revision),
                 Err(err) => {
                     counter!(METRIC_CACHE_BACKUP_REVISION_FAILURE_TOTAL).increment(1);
@@ -336,33 +336,6 @@ impl DataUsageCache {
         }
     }
 
-    async fn revision_for_path<S: ScannerObjectIO>(store: Arc<S>, path: &str) -> StorageResult<DataUsageCacheRevision> {
-        match store
-            .get_object_reader(
-                RUSTFS_META_BUCKET,
-                path,
-                None,
-                HeaderMap::new(),
-                &ObjectOptions {
-                    no_lock: true,
-                    ..Default::default()
-                },
-            )
-            .await
-        {
-            Ok(reader) => reader
-                .object_info
-                .etag
-                .filter(|etag| !etag.is_empty())
-                .map(DataUsageCacheRevision::Etag)
-                .ok_or_else(|| StorageError::other(format!("scanner cache object {path} has no ETag"))),
-            Err(Error::FileNotFound | Error::VolumeNotFound | Error::ObjectNotFound(_, _) | Error::BucketNotFound(_)) => {
-                Ok(DataUsageCacheRevision::Missing)
-            }
-            Err(err) => Err(err),
-        }
-    }
-
     pub(super) fn cache_save_timeout() -> Duration {
         crate::runtime_config::scanner_cache_save_timeout()
     }

crates/scanner/src/lib.rs

@@ -75,7 +75,10 @@ pub use remote_scanner::{
 };
 pub use runtime_config::{apply_scanner_runtime_config, scanner_runtime_config_status, validate_scanner_runtime_config};
 pub use rustfs_common::last_minute;
-pub use scanner::{ScannerCycleScheduleStatus, init_data_scanner, scanner_cycle_schedule_status, scanner_topology_digest};
+pub use scanner::{
+    ScannerCycleRecoveryMarker, ScannerCycleRecoveryStatus, ScannerCycleScheduleStatus, init_data_scanner,
+    reset_scanner_cycle_recovery, scanner_cycle_recovery_status, scanner_cycle_schedule_status, scanner_topology_digest,
+};
 pub use scanner_io::{
     ScannerDirtyUsageAckError, ScannerDirtyUsageState, acknowledge_dirty_usage_generation, clear_dirty_usage_bucket,
     record_dirty_usage_bucket, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_state,

crates/scanner/src/scanner.rs

@@ -20,7 +20,7 @@ use std::sync::{Arc, LazyLock, RwLock};
 
 use crate::data_usage_define::{
     BACKGROUND_HEAL_INFO_PATH, DATA_USAGE_BLOOM_NAME_PATH, DATA_USAGE_OBJ_NAME_PATH, DATA_USAGE_OBSERVED_OBJ_NAME_PATH,
-    DataUsageCache, DataUsageCacheRevision, LEGACY_DATA_USAGE_OBJ_NAME_PATH, read_config_with_revision,
+    DataUsageCache, DataUsageCacheRevision, LEGACY_DATA_USAGE_OBJ_NAME_PATH, read_config_revision, read_config_with_revision,
 };
 use crate::runtime_config::{
     ScannerRuntimeConfig, ScannerRuntimeConfigSource, refresh_scanner_runtime_config_from_global, scanner_bitrot_cycle,
@@ -54,9 +54,7 @@ use rustfs_config::{ENV_SCANNER_CYCLE, ENV_SCANNER_SPEED, ENV_SCANNER_START_DELA
 use rustfs_data_usage::observed_data_usage_is_newer;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest as _, Sha256};
-#[cfg(test)]
-use tokio::sync::Notify;
-use tokio::sync::mpsc;
+use tokio::sync::{Notify, mpsc};
 use tokio::time::{Duration, Instant};
 use tokio_util::sync::CancellationToken;
 use tokio_util::task::AbortOnDropHandle;
@@ -104,6 +102,13 @@ const CLEAN_IDLE_BACKOFF_FACTOR: u32 = 2;
 /// unavailable peer cannot drive a tight retry loop.
 const SCANNER_RETRY_BASE_INTERVAL: Duration = Duration::from_secs(5);
 const SCANNER_RETRY_MAX_INTERVAL: Duration = Duration::from_secs(30 * 60);
+/// A transient backend outage remains self-healing after the short retry
+/// budget is exhausted, but the probe is intentionally sparse until storage
+/// recovers or an operator reset wakes the scanner.
+const SCANNER_CYCLE_RECOVERY_PAUSED_INTERVAL: Duration = Duration::from_secs(5 * 60);
+/// Permanent recovery states still get a sparse status probe so a reset that
+/// races the wait registration cannot leave the scanner asleep forever.
+const SCANNER_CYCLE_RECOVERY_BLOCKED_PROBE_INTERVAL: Duration = Duration::from_secs(5 * 60);
 const SCANNER_LEADER_LOCK_POLL_INTERVAL: Duration = Duration::from_secs(1);
 #[cfg(not(test))]
 const SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT: Duration = Duration::from_secs(30);
@@ -125,6 +130,12 @@ type ScannerCycleStatePersistTestHook = (u64, Arc<Notify>);
 static SCANNER_CYCLE_STATE_PERSIST_TEST_HOOK: LazyLock<StdMutex<Option<ScannerCycleStatePersistTestHook>>> =
     LazyLock::new(|| StdMutex::new(None));
 
+static SCANNER_CYCLE_RECOVERY_WAKE: LazyLock<Notify> = LazyLock::new(Notify::new);
+
+pub(super) fn notify_scanner_cycle_recovery_wake() {
+    SCANNER_CYCLE_RECOVERY_WAKE.notify_one();
+}
+
 #[cfg(test)]
 struct ScannerCycleStatePersistTestHookGuard;
 
@@ -576,19 +587,21 @@ pub async fn init_data_scanner(ctx: CancellationToken, storeapi: Arc<ECStore>) {
             tokio::time::sleep(sleep_time).await;
         }
 
+        let mut transient_backoff = ScannerRetryBackoff::default();
+        let mut recovery_retry_count = 0_u32;
         loop {
             if ctx_clone.is_cancelled() {
                 break;
             }
 
-            if let Err(e) = run_data_scanner_with_maintenance_state(
+            let run_result = run_data_scanner_with_maintenance_state(
                 ctx_clone.clone(),
                 storeapi_clone.clone(),
                 startup_features,
                 startup_maintenance_generation,
             )
-            .await
-            {
+            .await;
+            if let Err(e) = &run_result {
                 error!(
                     target: "rustfs::scanner",
                     event = EVENT_SCANNER_CYCLE_STATE,
@@ -599,11 +612,52 @@ pub async fn init_data_scanner(ctx: CancellationToken, storeapi: Arc<ECStore>) {
                     "Scanner runtime iteration failed"
                 );
             }
+            let recovery_status = scanner_cycle_recovery_status();
+            if recovery_status.retryable {
+                recovery_retry_count = recovery_retry_count.saturating_add(1);
+                let _ = record_scanner_cycle_recovery_retry(recovery_retry_count);
+            } else {
+                recovery_retry_count = 0;
+            }
+
+            let recovery_status = scanner_cycle_recovery_status();
+            if recovery_status.state == "paused" {
+                transient_backoff.record_retryable_cycle(false);
+                tokio::select! {
+                    _ = ctx_clone.cancelled() => break,
+                    _ = SCANNER_CYCLE_RECOVERY_WAKE.notified() => {},
+                    _ = tokio::time::sleep(SCANNER_CYCLE_RECOVERY_PAUSED_INTERVAL) => {},
+                }
+                recovery_retry_count = 0;
+                continue;
+            }
+            if !recovery_status.retryable
+                && matches!(recovery_status.state.as_str(), "blocked" | "recovery-required" | "cleanup-pending")
+            {
+                transient_backoff.record_retryable_cycle(false);
+                tokio::select! {
+                    _ = ctx_clone.cancelled() => break,
+                    _ = SCANNER_CYCLE_RECOVERY_WAKE.notified() => {},
+                    _ = tokio::time::sleep(SCANNER_CYCLE_RECOVERY_BLOCKED_PROBE_INTERVAL) => {},
+                }
+                continue;
+            }
+
+            let retry_delay = if recovery_status.retryable || run_result.is_err() {
+                transient_backoff.record_retryable_cycle(true);
+                transient_backoff
+                    .retry_interval(scanner_cycle_interval())
+                    .unwrap_or(SCANNER_RETRY_BASE_INTERVAL)
+            } else {
+                transient_backoff.record_retryable_cycle(false);
+                randomized_cycle_delay()
+            };
             // Backoff before retrying after lock contention or scanner-level failures.
             // Keep this cancellation-aware so shutdown is not delayed by backoff sleep.
             tokio::select! {
                 _ = ctx_clone.cancelled() => break,
-                _ = tokio::time::sleep(randomized_cycle_delay()) => {}
+                _ = SCANNER_CYCLE_RECOVERY_WAKE.notified() => {},
+                _ = tokio::time::sleep(retry_delay) => {}
             }
         }
     });
@@ -1606,40 +1660,22 @@ async fn run_data_scanner_with_maintenance_state(
         observe_scanner_activity(&storeapi, distributed, &mut scanner_activity_seen).await;
     }
 
-    let (buf, mut cycle_revision) = match read_config_with_revision(storeapi.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str()).await {
-        Ok((buf, revision)) => (buf.unwrap_or_default(), revision),
-        Err(err) => {
-            error!(
-                target: "rustfs::scanner",
-                event = EVENT_SCANNER_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_RUNTIME,
-                path = %&*DATA_USAGE_BLOOM_NAME_PATH,
-                state = "revision_load_failed",
-                error = %err,
-                "Scanner cycle state revision load failed"
-            );
-            global_metrics().set_cycle(None).await;
-            return Ok(());
-        }
-    };
-    let (mut cycle_info, mut leader_epoch) = match decode_scanner_cycle_state_for_startup(&buf) {
-        Ok(state) => state,
-        Err(err) => {
-            error!(
-                target: "rustfs::scanner",
-                event = EVENT_SCANNER_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_RUNTIME,
-                path = %&*DATA_USAGE_BLOOM_NAME_PATH,
-                state = "cycle_decode_failed",
-                error = %err,
-                "Scanner stopped because persisted cycle state is invalid"
-            );
-            global_metrics().set_cycle(None).await;
-            return Ok(());
-        }
-    };
+    let (mut cycle_info, mut leader_epoch, mut cycle_revision) =
+        match load_scanner_cycle_state_for_startup(storeapi.clone()).await {
+            ScannerCycleStateStartup::Ready {
+                cycle,
+                leader_epoch,
+                revision,
+            } => (cycle, leader_epoch, revision),
+            ScannerCycleStateStartup::Blocked => {
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleStateStartup::Transient(err) => {
+                global_metrics().set_cycle(None).await;
+                return Err(err);
+            }
+        };
     let usage_floor = match persisted_usage_floor(storeapi.clone()).await {
         Ok(floor) => floor,
         Err(err) => {
@@ -2219,7 +2255,12 @@ pub(crate) use activity::{
 pub(crate) use activity::{ScannerCycleOutcome, scanner_cycle_outcome_with_pending_maintenance};
 #[cfg(test)]
 pub(crate) use cycle_state::encode_scanner_cycle_fence_for_test;
-pub(crate) use cycle_state::{current_scanner_leader_epoch, decode_persisted_scanner_cycle_fence};
+pub use cycle_state::{
+    ScannerCycleRecoveryMarker, ScannerCycleRecoveryStatus, reset_scanner_cycle_recovery, scanner_cycle_recovery_status,
+};
+pub(crate) use cycle_state::{
+    current_scanner_leader_epoch, decode_persisted_scanner_cycle_fence, load_scanner_cycle_state_for_startup,
+};
 pub use heal_info::{BackgroundHealInfo, read_background_heal_info, save_background_heal_info};
 pub use usage_store::store_data_usage_in_backend;
 

crates/scanner/src/scanner/leadership.rs

@@ -196,7 +196,7 @@ pub(super) async fn claim_scanner_leadership(
         if ctx.is_cancelled() {
             return false;
         }
-        let Some(claimed_epoch) = persisted_epoch.checked_add(1) else {
+        let Some(claimed_epoch) = persisted_epoch.checked_add(1).filter(|epoch| *epoch < u64::MAX) else {
             error!(
                 target: "rustfs::scanner",
                 event = EVENT_SCANNER_PERSIST_STATE,

crates/scanner/src/scanner/tests.rs

@@ -15,11 +15,12 @@
 use super::*;
 use crate::EcstoreResult;
 use crate::{
-    Endpoint, EndpointServerPools, Endpoints, InstanceContext, PoolEndpoints, ScannerGetObjectReader as GetObjectReader,
-    ScannerObjectInfo as ObjectInfo, ScannerObjectOptions as ObjectOptions, ScannerPutObjReader as PutObjReader,
-    init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests, init_local_disks_with_instance_ctx,
+    DATA_USAGE_BLOOM_RECOVERY_PATH, Endpoint, EndpointServerPools, Endpoints, InstanceContext, PoolEndpoints,
+    ScannerGetObjectReader as GetObjectReader, ScannerObjectInfo as ObjectInfo, ScannerObjectOptions as ObjectOptions,
+    ScannerPutObjReader as PutObjReader, init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests,
+    init_local_disks_with_instance_ctx,
 };
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::io::Cursor;
 use std::task::Poll;
 use temp_env::{with_var, with_var_unset};
@@ -117,6 +118,15 @@ async fn scanner_cycle_lock_fence_bounds_uncooperative_shutdown() {
     assert!(cycle_ctx.is_cancelled());
 }
 
+#[tokio::test]
+async fn scanner_cycle_recovery_wake_survives_wait_registration_race() {
+    notify_scanner_cycle_recovery_wake();
+
+    tokio::time::timeout(Duration::from_secs(1), SCANNER_CYCLE_RECOVERY_WAKE.notified())
+        .await
+        .expect("recovery wake should retain a permit until the waiter registers");
+}
+
 struct ScannerDefaultSpeedGuard;
 
 impl ScannerDefaultSpeedGuard {
@@ -151,6 +161,7 @@ impl Drop for ScannerDefaultCycleGuard {
 struct MemoryConfigStore {
     objects: Mutex<HashMap<String, Vec<u8>>>,
     revisions: Mutex<HashMap<String, u64>>,
+    non_regular_objects: Mutex<HashSet<String>>,
     fail_put_number: Mutex<HashMap<String, usize>>,
     object_not_found_put_number: Mutex<HashMap<String, usize>>,
     error_after_commit_put_number: Mutex<HashMap<String, usize>>,
@@ -191,12 +202,16 @@ impl crate::storage_api::scanner_io::ObjectIO for MemoryConfigStore {
             .get(&key)
             .cloned()
             .ok_or(EcstoreError::FileNotFound)?;
-        let revision = *self.revisions.lock().await.entry(key).or_insert(1);
+        let data_len = i64::try_from(data.len()).expect("memory test object length should fit in i64");
+        let revision = *self.revisions.lock().await.entry(key.clone()).or_insert(1);
+        let is_dir = self.non_regular_objects.lock().await.contains(&key);
 
         Ok(GetObjectReader {
             stream: Box::new(Cursor::new(data)),
             object_info: ObjectInfo {
                 etag: Some(format!("memory-{revision}")),
+                size: data_len,
+                is_dir,
                 ..Default::default()
             },
             buffered_body: None,
@@ -797,6 +812,10 @@ fn scanner_cycle_state_decodes_legacy_and_fenced_formats() {
     let (fenced_cycle, fenced_epoch) = decode_scanner_cycle_state(&fenced).expect("fenced cycle state should decode");
     assert_eq!(fenced_cycle.next, 13);
     assert_eq!(fenced_epoch, 7);
+
+    let mut trailing = fenced;
+    trailing.push(0);
+    assert!(decode_scanner_cycle_state(&trailing).is_err());
 }
 
 #[test]
@@ -823,6 +842,840 @@ fn scanner_startup_fails_closed_on_nonempty_corrupt_cycle_state() {
     assert!(encode_scanner_cycle_state(&exhausted, 7).is_err());
 }
 
+#[tokio::test]
+async fn corrupt_cycle_state_is_quarantined_once() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.objects.lock().await.insert(state_key.clone(), vec![1]);
+    store.revisions.lock().await.insert(state_key.clone(), 7);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store.clone()).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+    let marker_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    let marker_data = store
+        .objects
+        .lock()
+        .await
+        .get(&marker_key)
+        .cloned()
+        .expect("corrupt state must leave a durable recovery marker");
+    let marker: ScannerCycleRecoveryMarker = serde_json::from_slice(&marker_data).expect("marker should be valid JSON");
+    assert_eq!(marker.primary_revision, "memory-7");
+    assert_eq!(marker.path, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    assert_eq!(marker.quarantine_path, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    assert_eq!(marker.classification, "corrupt");
+
+    // A second startup sees the matching marker before consuming the poison body.
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store.clone()).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+
+    // Replacing the primary object advances its revision; the stale marker must
+    // not quarantine the newer, valid state.
+    let cycle = CurrentCycle {
+        next: 9,
+        ..Default::default()
+    };
+    let encoded = encode_scanner_cycle_state(&cycle, 3).expect("valid state should encode");
+    store.objects.lock().await.insert(state_key.clone(), encoded);
+    store.revisions.lock().await.insert(state_key, 8);
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Ready {
+            cycle: CurrentCycle { next: 9, .. },
+            leader_epoch: 3,
+            ..
+        }
+    ));
+}
+
+#[tokio::test]
+async fn empty_cycle_state_object_is_quarantined_as_corrupt() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.objects.lock().await.insert(state_key.clone(), Vec::new());
+    store.revisions.lock().await.insert(state_key, 6);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+    assert_eq!(scanner_cycle_recovery_status().classification.as_deref(), Some("corrupt"));
+    assert!(
+        scanner_cycle_recovery_status()
+            .reason
+            .as_deref()
+            .is_some_and(|reason| reason.contains("empty"))
+    );
+}
+
+#[tokio::test]
+async fn future_cycle_state_schema_is_recovery_required() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    let mut future = 17_u64.to_le_bytes().to_vec();
+    future.extend_from_slice(b"RSCYC999");
+    future.extend_from_slice(&4_u64.to_le_bytes());
+    future.extend_from_slice(&[0x90]);
+    store.objects.lock().await.insert(state_key.clone(), future);
+    store.revisions.lock().await.insert(state_key, 13);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+    assert_eq!(scanner_cycle_recovery_status().classification.as_deref(), Some("future_schema"));
+}
+
+#[tokio::test]
+async fn concurrent_leaders_cannot_quarantine_newer_cycle_state() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.objects.lock().await.insert(state_key.clone(), vec![1]);
+    store.revisions.lock().await.insert(state_key, 4);
+
+    let (first, second) = tokio::join!(
+        load_scanner_cycle_state_for_startup(store.clone()),
+        load_scanner_cycle_state_for_startup(store.clone()),
+    );
+    assert!(matches!(first, ScannerCycleStateStartup::Blocked));
+    assert!(matches!(second, ScannerCycleStateStartup::Blocked));
+
+    let marker_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    let marker_data = store
+        .objects
+        .lock()
+        .await
+        .get(&marker_key)
+        .cloned()
+        .expect("one contender must publish the recovery marker");
+    let marker: ScannerCycleRecoveryMarker = serde_json::from_slice(&marker_data).expect("marker should decode");
+    assert_eq!(marker.primary_revision, "memory-4");
+}
+
+#[tokio::test]
+async fn cleanup_pending_marker_blocks_a_rewritten_primary_after_restart() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    let marker_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    let encoded = encode_scanner_cycle_state(
+        &CurrentCycle {
+            next: 12,
+            ..Default::default()
+        },
+        8,
+    )
+    .expect("valid state should encode");
+    store.objects.lock().await.insert(state_key.clone(), encoded);
+    store.revisions.lock().await.insert(state_key, 22);
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: "memory-21".to_string(),
+        generation: 11,
+        leader_epoch: 7,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 1,
+        reason: "reset in progress".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: "cleanup-pending".to_string(),
+    };
+    store
+        .objects
+        .lock()
+        .await
+        .insert(marker_key.clone(), serde_json::to_vec(&marker).expect("marker should encode"));
+    store.revisions.lock().await.insert(marker_key, 3);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+    assert_eq!(scanner_cycle_recovery_status().state, "cleanup-pending");
+}
+
+#[test]
+fn full_rescan_reset_accepts_unknown_marker_fields_without_trusting_cursor() {
+    let marker = br#"{
+        "schema_version": 99,
+        "primary_revision": "memory-7",
+        "generation": 9000,
+        "leader_epoch": 9000,
+        "classification": "new-future-classification",
+        "first_detected_at_unix_secs": 1,
+        "last_attempt_at_unix_secs": 2,
+        "retry_count": 9,
+        "reason": "future marker",
+        "path": "buckets/.bloomcycle.bin",
+        "quarantine_path": "buckets/.bloomcycle.bin.recovery-required.json",
+        "future_field": {"cursor": "untrusted"}
+    }"#;
+    let decoded =
+        super::cycle_state::decode_recovery_marker_for_reset(marker, &DataUsageCacheRevision::Etag("memory-3".to_string()))
+            .expect("full-rescan compatibility decoder should accept additive fields");
+    assert_eq!(decoded.primary_revision, "memory-7");
+    assert_eq!(decoded.classification, "future_schema");
+    assert_eq!(decoded.generation, 0);
+    assert_eq!(decoded.leader_epoch, 0);
+    assert_eq!(decoded.state, "blocked");
+
+    let malformed =
+        super::cycle_state::decode_recovery_marker_for_reset(b"{not-json", &DataUsageCacheRevision::Etag("memory-4".to_string()))
+            .expect("a full-rescan reset must recover even when the marker is malformed");
+    assert!(malformed.primary_revision.is_empty());
+    assert_eq!(malformed.classification, "future_schema");
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_after_malformed_marker_without_trusting_cursor() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0xff, 0x00, 0x01])
+        .await
+        .expect("corrupt cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), br#"{not-json"#.to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("full-rescan reset should recover malformed marker");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(cycle.next, 0, "reset must use the verified usage floor, not marker cursor");
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_ignores_epoch_from_malformed_future_primary() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let mut future_primary = vec![0; 24];
+    future_primary[8..16].copy_from_slice(b"RSCY9999");
+    future_primary[16..24].copy_from_slice(&u64::MAX.to_le_bytes());
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), future_primary)
+        .await
+        .expect("future cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), br#"{not-json"#.to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("full-rescan reset should recover malformed future state");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (_, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(leader_epoch, 1, "invalid persisted bytes must not raise the recovery epoch");
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn ecstore_exact_recovery_marker_delete_honors_etag() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"marker-v1".to_vec())
+        .await
+        .expect("initial recovery marker should be persisted");
+    let (_, stale_revision) = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+        .await
+        .expect("initial marker revision should load");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"marker-v2".to_vec())
+        .await
+        .expect("replacement recovery marker should be persisted");
+
+    let delete_result = store
+        .delete_config_object(
+            RUSTFS_META_BUCKET,
+            DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+            ObjectOptions {
+                http_preconditions: Some(stale_revision.preconditions()),
+                ..Default::default()
+            },
+        )
+        .await;
+    assert!(matches!(delete_result, Err(EcstoreError::PreconditionFailed)));
+    assert_eq!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+            .await
+            .expect("replacement marker should remain durable"),
+        b"marker-v2"
+    );
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rejects_corrupt_primary_under_stale_blocked_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let corrupt_primary = vec![0xff, 0x00, 0x01];
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), corrupt_primary.clone())
+        .await
+        .expect("corrupt cycle state should be persisted");
+    let (_, primary_revision) = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("primary revision should load");
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: "memory-stale".to_string(),
+        generation: 1,
+        leader_epoch: 1,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 1,
+        reason: "blocked primary changed".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: "blocked".to_string(),
+    };
+    let marker_data = serde_json::to_vec(&marker).expect("blocked marker should encode");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), marker_data.clone())
+        .await
+        .expect("blocked marker should be persisted");
+
+    assert!(
+        reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+            .await
+            .is_err(),
+        "a strict marker must fail closed when its primary revision changed"
+    );
+    assert_eq!(
+        read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+            .await
+            .expect("primary should remain readable"),
+        corrupt_primary
+    );
+    assert_eq!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+            .await
+            .expect("blocked marker should remain durable"),
+        marker_data
+    );
+    assert!(!matches!(primary_revision, DataUsageCacheRevision::Missing));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_preserves_valid_primary_when_marker_is_malformed() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let primary = CurrentCycle {
+        next: 42,
+        ..Default::default()
+    };
+    let old_primary_data = encode_scanner_cycle_state(&primary, 7).expect("valid cycle state should encode");
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), old_primary_data.clone())
+        .await
+        .expect("valid cycle state should be persisted");
+    let (_, old_primary_revision) = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("primary state revision should load");
+    let old_usage = DataUsageInfo {
+        scanner_epoch: Some(7),
+        scanner_cycle: Some(41),
+        ..Default::default()
+    };
+    let old_usage_data = serde_json::to_vec(&old_usage).expect("usage snapshot should encode");
+    save_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str(), old_usage_data.clone())
+        .await
+        .expect("usage snapshot should be persisted");
+    let (_, old_usage_revision) = read_config_with_revision(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("usage snapshot revision should load");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("reset should clear a stale malformed marker");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("valid primary should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("primary cycle state should decode");
+    assert_eq!(cycle.next, 42, "reset must not regress an independently fenced primary");
+    assert_eq!(leader_epoch, 8, "reset must advance the preserved primary epoch");
+    let stale_primary_save = save_config_with_preconditions(
+        store.clone(),
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        old_primary_data,
+        old_primary_revision.preconditions(),
+    )
+    .await;
+    assert!(matches!(stale_primary_save, Err(EcstoreError::PreconditionFailed)));
+    let usage = read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("usage epoch fence should remain durable");
+    assert_eq!(
+        serde_json::from_slice::<DataUsageInfo>(&usage)
+            .expect("fenced usage should decode")
+            .scanner_epoch,
+        Some(8)
+    );
+    let stale_save = save_config_with_preconditions(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        old_usage_data,
+        old_usage_revision.preconditions(),
+    )
+    .await;
+    assert!(matches!(stale_save, Err(EcstoreError::PreconditionFailed)));
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_resumes_cleanup_pending_preserved_primary() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let completed_at = Utc::now();
+    let primary = CurrentCycle {
+        current: 3,
+        next: 42,
+        cycle_completed: vec![completed_at],
+        started: completed_at,
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        encode_scanner_cycle_state(&primary, 7).expect("valid cycle state should encode"),
+    )
+    .await
+    .expect("valid cycle state should be persisted");
+    let usage = DataUsageInfo {
+        scanner_epoch: Some(7),
+        scanner_cycle: Some(41),
+        ..Default::default()
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&usage).expect("usage snapshot should encode"),
+    )
+    .await
+    .expect("usage snapshot should be persisted");
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: "memory-old".to_string(),
+        generation: 41,
+        leader_epoch: 7,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 1,
+        reason: "reset in progress".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: "cleanup-pending".to_string(),
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        serde_json::to_vec(&marker).expect("marker should encode"),
+    )
+    .await
+    .expect("cleanup marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("reset should resume a cleanup-pending preserved primary");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("preserved cycle state should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("cycle state should decode");
+    assert_eq!(cycle.current, 3, "cleanup retry must preserve the in-progress cursor");
+    assert_eq!(cycle.next, 42);
+    assert_eq!(cycle.cycle_completed, vec![completed_at]);
+    assert_eq!(cycle.started, completed_at);
+    assert_eq!(leader_epoch, 8);
+    let usage = read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("usage epoch fence should remain durable");
+    assert_eq!(
+        serde_json::from_slice::<DataUsageInfo>(&usage)
+            .expect("usage should decode")
+            .scanner_epoch,
+        Some(8)
+    );
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_oversized_regular_primary_with_malformed_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0; 1024 * 1024 + 1])
+        .await
+        .expect("oversized cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("explicit full-rescan reset should replace an oversized regular primary");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(cycle.next, 0);
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_oversized_primary_after_cleanup_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0; 1024 * 1024 + 1])
+        .await
+        .expect("oversized cycle state should be persisted");
+    let (_, primary_revision) = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("primary revision should load");
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: match primary_revision {
+            DataUsageCacheRevision::Etag(etag) => etag,
+            DataUsageCacheRevision::Missing => panic!("primary revision should be present"),
+        },
+        generation: 1,
+        leader_epoch: 1,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 1,
+        reason: "reset in progress".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: "cleanup-pending".to_string(),
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        serde_json::to_vec(&marker).expect("cleanup marker should encode"),
+    )
+    .await
+    .expect("cleanup marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("cleanup retry should rebuild an oversized primary");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(cycle.next, 0);
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_with_oversized_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0xff, 0x00, 0x01])
+        .await
+        .expect("corrupt cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), vec![b'x'; 64 * 1024 + 1])
+        .await
+        .expect("oversized recovery marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("full-rescan reset should recover an oversized marker");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (_, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_with_empty_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0xff, 0x00, 0x01])
+        .await
+        .expect("corrupt cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), Vec::new())
+        .await
+        .expect("empty recovery marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("full-rescan reset should recover an empty marker");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (_, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_keeps_cleanup_marker_when_preserved_epoch_is_exhausted() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let primary = CurrentCycle {
+        next: 42,
+        ..Default::default()
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        encode_scanner_cycle_state(&primary, u64::MAX).expect("valid cycle state should encode"),
+    )
+    .await
+    .expect("valid cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    assert!(
+        reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+            .await
+            .is_err()
+    );
+
+    let marker = read_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+        .await
+        .expect("cleanup marker should remain durable");
+    assert_eq!(
+        serde_json::from_slice::<ScannerCycleRecoveryMarker>(&marker)
+            .expect("cleanup marker should decode")
+            .state,
+        "cleanup-pending"
+    );
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rejects_preserved_epoch_that_would_be_terminal() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let primary = CurrentCycle {
+        next: 42,
+        ..Default::default()
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        encode_scanner_cycle_state(&primary, u64::MAX - 1).expect("valid cycle state should encode"),
+    )
+    .await
+    .expect("valid cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    assert!(
+        reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+            .await
+            .is_err(),
+        "reset must not persist the terminal leader epoch"
+    );
+
+    let marker = read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+        .await
+        .expect("cleanup marker should remain durable");
+    assert_eq!(
+        serde_json::from_slice::<ScannerCycleRecoveryMarker>(&marker)
+            .expect("cleanup marker should decode")
+            .state,
+        "cleanup-pending"
+    );
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rejects_usage_floor_that_would_be_terminal() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), vec![0xff, 0x00, 0x01])
+        .await
+        .expect("corrupt cycle state should be persisted");
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&DataUsageInfo {
+            scanner_epoch: Some(u64::MAX - 1),
+            ..Default::default()
+        })
+        .expect("usage floor should encode"),
+    )
+    .await
+    .expect("usage floor should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    assert!(
+        reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+            .await
+            .is_err(),
+        "reset must not persist the terminal leader epoch"
+    );
+    assert_eq!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+            .await
+            .expect("recovery marker should remain durable"),
+        b"{not-json"
+    );
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_empty_primary_with_malformed_marker() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), Vec::new())
+        .await
+        .expect("empty cycle state should be persisted");
+    save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), b"{not-json".to_vec())
+        .await
+        .expect("malformed marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("explicit full-rescan reset should replace an empty primary");
+
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle state should remain durable");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(cycle.next, 0);
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn full_rescan_reset_rebuilds_when_primary_cycle_state_is_missing() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: "memory-missing".to_string(),
+        generation: u64::MAX,
+        leader_epoch: u64::MAX,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 0,
+        reason: "missing primary".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: "blocked".to_string(),
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        serde_json::to_vec(&marker).expect("marker should encode"),
+    )
+    .await
+    .expect("marker should be persisted");
+
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("full-rescan reset should recreate missing primary");
+    let state = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("missing primary should be rebuilt");
+    let (cycle, leader_epoch) = decode_scanner_cycle_state(&state).expect("rebuilt cycle state should decode");
+    assert_eq!(cycle.next, 0);
+    assert_eq!(leader_epoch, 1);
+    assert!(matches!(
+        read_config(store, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[tokio::test]
+async fn corrupt_cycle_state_rename_or_marker_failure_stays_recovery_required() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let state_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    let marker_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    store.objects.lock().await.insert(state_key.clone(), vec![1]);
+    store.revisions.lock().await.insert(state_key, 9);
+    store.fail_put_number.lock().await.insert(marker_key, 1);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store.clone()).await,
+        ScannerCycleStateStartup::Transient(_)
+    ));
+    let status = scanner_cycle_recovery_status();
+    assert_eq!(status.state, "recovery-required");
+    assert!(status.retryable);
+    assert!(
+        store
+            .objects
+            .lock()
+            .await
+            .contains_key(&memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str()))
+    );
+}
+
+#[tokio::test]
+async fn oversized_or_symlinked_cycle_state_is_rejected() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.objects.lock().await.insert(key.clone(), vec![0; 1024 * 1024 + 1]);
+    store.revisions.lock().await.insert(key.clone(), 11);
+
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store.clone()).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+    assert_eq!(scanner_cycle_recovery_status().classification.as_deref(), Some("corrupt"));
+    assert!(
+        scanner_cycle_recovery_status()
+            .reason
+            .as_deref()
+            .is_some_and(|reason| reason.contains("oversized"))
+    );
+
+    let marker_key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str());
+    store.objects.lock().await.remove(&marker_key);
+    store.objects.lock().await.insert(key.clone(), vec![1]);
+    store.revisions.lock().await.insert(key.clone(), 12);
+    store.non_regular_objects.lock().await.insert(key);
+    // The object contract exposes a non-regular object as `is_dir`; local
+    // backends reject symlink/reparse entries before they become an object.
+    assert!(matches!(
+        load_scanner_cycle_state_for_startup(store).await,
+        ScannerCycleStateStartup::Blocked
+    ));
+}
+
 #[tokio::test]
 async fn scanner_startup_uses_primary_and_backup_usage_floor() {
     let store = Arc::new(MemoryConfigStore::default());
@@ -855,6 +1708,31 @@ async fn scanner_startup_uses_primary_and_backup_usage_floor() {
     assert_eq!(epoch, 11);
 }
 
+#[tokio::test]
+async fn scanner_usage_floor_ignores_older_backup_after_primary_epoch_fence() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let backup_path = format!("{}.bkp", DATA_USAGE_OBJ_NAME_PATH.as_str());
+    for (path, epoch, cycle) in [(DATA_USAGE_OBJ_NAME_PATH.as_str(), 8, 100), (backup_path.as_str(), 7, 10_000)] {
+        store.objects.lock().await.insert(
+            memory_config_key(RUSTFS_META_BUCKET, path),
+            serde_json::to_vec(&DataUsageInfo {
+                scanner_epoch: Some(epoch),
+                scanner_cycle: Some(cycle),
+                ..Default::default()
+            })
+            .expect("usage snapshot should encode"),
+        );
+    }
+
+    assert_eq!(
+        persisted_usage_floor(store).await.expect("usage floor should load"),
+        PersistedUsageFloor {
+            next_cycle: 101,
+            leader_epoch: 8,
+        }
+    );
+}
+
 #[test]
 fn scanner_startup_treats_incomplete_usage_snapshot_as_cold() {
     let mut legacy = complete_usage_with_bucket_count(Some(std::time::SystemTime::now()), 1);
@@ -987,6 +1865,15 @@ async fn scanner_usage_floor_fails_closed_on_corrupt_or_exhausted_usage_state()
 
     assert!(persisted_usage_floor(store.clone()).await.is_err());
 
+    store.objects.lock().await.insert(
+        memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_OBJ_NAME_PATH.as_str()),
+        br#"{}"#.to_vec(),
+    );
+    assert!(
+        persisted_usage_floor(store.clone()).await.is_err(),
+        "a structurally incomplete usage snapshot must not be treated as an empty floor"
+    );
+
     store.objects.lock().await.insert(
         memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_OBJ_NAME_PATH.as_str()),
         serde_json::to_vec(&DataUsageInfo {
@@ -1244,6 +2131,22 @@ async fn test_leadership_claim_preserves_usage_epoch_floor_across_old_epoch_conf
     assert_eq!(store.put_counts.lock().await.get(&key), Some(&3));
 }
 
+#[tokio::test]
+async fn test_leadership_claim_rejects_terminal_epoch() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let ctx = CancellationToken::new();
+    let mut revision = DataUsageCacheRevision::Missing;
+    let mut cycle = CurrentCycle {
+        next: 12,
+        ..Default::default()
+    };
+    let mut persisted_epoch = u64::MAX - 1;
+
+    assert!(!claim_scanner_leadership(&ctx, store.clone(), &mut cycle, &mut revision, &mut persisted_epoch).await);
+    assert_eq!(persisted_epoch, u64::MAX - 1);
+    assert!(read_config(store, &DATA_USAGE_BLOOM_NAME_PATH).await.is_err());
+}
+
 #[tokio::test]
 async fn test_leadership_claim_confirms_commit_after_returned_error() {
     let store = Arc::new(MemoryConfigStore::default());
@@ -2975,6 +3878,24 @@ fn superseded_retry_backoff_grows_from_the_default_cycle() {
     }
 }
 
+#[tokio::test(start_paused = true)]
+async fn corrupt_cycle_state_backoff_uses_virtual_clock() {
+    let mut backoff = ScannerRetryBackoff::default();
+    backoff.record_retryable_cycle(true);
+    let first_delay = backoff
+        .retry_interval(Duration::from_secs(60))
+        .expect("the first recovery retry should be scheduled");
+    assert_eq!(first_delay, Duration::from_secs(5));
+
+    let deadline = Instant::now() + first_delay;
+    assert!(Instant::now() < deadline);
+    tokio::time::advance(first_delay).await;
+    assert!(Instant::now() >= deadline);
+
+    backoff.record_retryable_cycle(true);
+    assert_eq!(backoff.retry_interval(Duration::from_secs(60)), Some(Duration::from_secs(10)));
+}
+
 #[test]
 fn scanner_cycle_wait_plan_drives_growth_resets_and_bitrot_cap() {
     let runtime_config = ScannerRuntimeConfig {

crates/scanner/src/scanner_folder.rs

@@ -43,10 +43,7 @@ use rustfs_common::metrics::{
     UpdateCurrentPathFn, current_path_updater, global_metrics,
 };
 use rustfs_common::trace_bus::{TraceEvent, TraceFunc, TraceKind, trace_emit, trace_subscriber_count};
-use rustfs_filemeta::{
-    MAX_META_CACHE_HEAL_CANDIDATES, MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS, MetaCacheEntries, MetaCacheEntry,
-    MetaCacheHealCandidateKind,
-};
+use rustfs_filemeta::{MetaCacheEntries, MetaCacheEntry, MetadataResolutionParams};
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
 use s3s::dto::{BucketLifecycleConfiguration, ObjectLockConfiguration, VersioningConfiguration};
 use time::OffsetDateTime;
@@ -99,11 +96,6 @@ const METRIC_SCANNER_EXCESS_OBJECT_VERSION_SIZE_TOTAL: &str = "rustfs_scanner_ex
 const METRIC_SCANNER_EXCESS_FOLDERS_TOTAL: &str = "rustfs_scanner_excess_folders_total";
 const METRIC_SCANNER_PENDING_HEAL_PRUNE_TOTAL: &str = "rustfs_scanner_pending_heal_prune_total";
 const METRIC_SCANNER_PENDING_HEAL_MALFORMED_TOTAL: &str = "rustfs_scanner_pending_heal_malformed_total";
-const METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL: &str = "rustfs_scanner_heal_discovery_candidates_total";
-const METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL: &str = "rustfs_scanner_heal_discovery_sub_quorum_total";
-const METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL: &str = "rustfs_scanner_heal_discovery_unverified_total";
-const METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL: &str = "rustfs_scanner_heal_discovery_queued_total";
-const METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL: &str = "rustfs_scanner_heal_discovery_truncated_total";
 const MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET: usize = 128;
 
 // --- scanner excess alerts as S3 notification events (rustfs/backlog#1868) --
@@ -891,7 +883,7 @@ impl FolderScanner {
         object: Option<String>,
         version_id: Option<String>,
         request: HealChannelRequest,
-    ) -> Result<HealAdmissionResult, ScannerError> {
+    ) -> Result<(), ScannerError> {
         let candidate_type = pending_scanner_heal_candidate_type(kind);
         let priority = request.priority;
         let scan_mode = request.scan_mode.unwrap_or(self.scan_mode);
@@ -919,7 +911,7 @@ impl FolderScanner {
                     error = %err,
                     "Scanner deferred heal request after channel error"
                 );
-                return Ok(HealAdmissionResult::Full);
+                return Ok(());
             }
         };
         self.update_pending_scanner_heal_after_admission(
@@ -931,7 +923,7 @@ impl FolderScanner {
             result,
         );
         if result.is_admitted() {
-            return Ok(result);
+            return Ok(());
         }
 
         record_high_priority_heal_escalation(candidate_type, priority, result);
@@ -952,7 +944,7 @@ impl FolderScanner {
             state = "high_priority_not_admitted",
             "Scanner high-priority heal admission failed"
         );
-        Ok(result)
+        Ok(())
     }
 
     pub fn set_heal_object_select(&mut self, prob: u32) {
@@ -1744,7 +1736,14 @@ impl FolderScanner {
                 break;
             }
 
-            let mut previous_bucket = String::new();
+            let mut resolver = MetadataResolutionParams {
+                dir_quorum: self.disks_quorum,
+                obj_quorum: self.disks_quorum,
+                bucket: "".to_string(),
+                strict: false,
+                ..Default::default()
+            };
+
             for name in abandoned_children {
                 if !self.should_heal().await {
                     break;
@@ -1752,7 +1751,7 @@ impl FolderScanner {
 
                 let (bucket, prefix) = path2_bucket_object(name.as_str());
 
-                if bucket != previous_bucket {
+                if bucket != resolver.bucket {
                     self.send_required_scanner_heal_request(
                         PendingScannerHealKind::Bucket,
                         bucket.clone(),
@@ -1761,9 +1760,10 @@ impl FolderScanner {
                         build_bucket_heal_request(bucket.clone(), HealChannelPriority::High),
                     )
                     .await?;
-                    previous_bucket = bucket.clone();
                 }
 
+                resolver.bucket = bucket.clone();
+
                 let child_ctx = ctx.child_token();
 
                 let (agreed_tx, mut agreed_rx) = mpsc::channel::<String>(1);
@@ -1880,8 +1880,6 @@ impl FolderScanner {
                 let mut agreed_closed = false;
                 let mut partial_closed = false;
                 let mut finished_closed = false;
-                let mut seen_heal_candidates: HashSet<(String, Option<String>, MetaCacheHealCandidateKind)> = HashSet::new();
-                let mut seen_truncated_objects: HashSet<String> = HashSet::new();
 
                 loop {
                     if agreed_closed && partial_closed && finished_closed {
@@ -1906,117 +1904,65 @@ impl FolderScanner {
                                 break;
                             }
 
-                            let discovery = entries.discover_heal_candidates(&bucket, MAX_META_CACHE_HEAL_CANDIDATES);
-                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL)
-                                .increment(u64::try_from(discovery.candidates.len()).unwrap_or(u64::MAX));
-                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL).increment(
-                                u64::try_from(
-                                    discovery
-                                        .candidates
-                                        .iter()
-                                        .filter(|candidate| candidate.replica_count < disks_quorum)
-                                        .count(),
-                                )
-                                .unwrap_or(u64::MAX),
-                            );
-                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL).increment(
-                                u64::try_from(discovery.unverified_count).unwrap_or(u64::MAX),
-                            );
-                            if discovery.truncated {
-                                counter!(METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL).increment(1);
+                            let Some(entry) = resolve_object_heal_entry(&entries, resolver.clone()) else {
+                                continue;
+                            };
+
+                            (self.update_current_path)(&entry.name).await;
+
+                            if entry.is_dir() {
+                                continue;
                             }
 
-                            for candidate in discovery.candidates {
-                                let sub_quorum_candidate = candidate.replica_count < disks_quorum;
-                                let version_id = candidate.validated_version().map(|id| id.to_string());
-                                let identity = (candidate.object.clone(), version_id.clone(), candidate.kind.clone());
-                                if seen_heal_candidates.len() >= MAX_META_CACHE_HEAL_CANDIDATES
-                                    && !seen_heal_candidates.contains(&identity)
-                                {
-                                    continue;
-                                }
-                                if !seen_heal_candidates.insert(identity) {
-                                    continue;
-                                }
-                                let request = if candidate.is_unversioned() {
-                                    build_non_destructive_object_heal_request(
-                                        bucket.clone(),
-                                        candidate.object.clone(),
-                                        self.scan_mode,
-                                        HealChannelPriority::High,
-                                    )
-                                } else {
-                                    build_object_heal_request(
-                                        bucket.clone(),
-                                        candidate.object.clone(),
-                                        version_id.clone(),
-                                        self.scan_mode,
-                                        HealChannelPriority::High,
-                                    )
-                                };
-                                (self.update_current_path)(&candidate.object).await;
-                                let admission = self.send_required_scanner_heal_request(
-                                    PendingScannerHealKind::Object,
-                                    bucket.clone(),
-                                    Some(candidate.object.clone()),
-                                    version_id.clone(),
-                                    request,
-                                )
-                                .await?;
-                                if admission.is_admitted() {
-                                    counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
-                                } else if sub_quorum_candidate {
-                                    self.mark_pending_scanner_heal_reason(
-                                        PendingScannerHealKind::Object,
-                                        &bucket,
-                                        Some(&candidate.object),
-                                        version_id.as_deref(),
-                                        "sub_quorum_metadata",
+                            let fivs = match entry.file_info_versions(&bucket) {
+                                Ok(fivs) => fivs,
+                                Err(e) => {
+                                    error!(
+                                        target: "rustfs::scanner::folder",
+                                        event = EVENT_SCANNER_FOLDER_STATE,
+                                        component = LOG_COMPONENT_SCANNER,
+                                        subsystem = LOG_SUBSYSTEM_FOLDER,
+                                        bucket = %bucket,
+                                        entry = %entry.name,
+                                        state = "file_info_versions_failed",
+                                        error = %e,
+                                        "Scanner list_path_raw failed to resolve file versions"
                                     );
-                                }
-                                found_objects = true;
-                            }
-
-                            // A bounded candidate union may overflow for an
-                            // object with a very long version history. Keep
-                            // that overflow explicit and issue one safe,
-                            // versionless inspection request per object so
-                            // the dropped versions are not silently treated
-                            // as absent. This continuation is deliberately
-                            // outside the versioned candidate cap and always
-                            // disables destructive cleanup.
-                            for object in discovery.truncated_objects {
-                                if seen_truncated_objects.len() >= MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
-                                    && !seen_truncated_objects.contains(&object)
-                                {
-                                    continue;
-                                }
-                                if !seen_truncated_objects.insert(object.clone()) {
-                                    continue;
-                                }
-                                let identity = (object.clone(), None, MetaCacheHealCandidateKind::UnversionedObject);
-                                if !seen_heal_candidates.insert(identity) {
-                                    continue;
-                                }
-                                let request = build_non_destructive_object_heal_request(
-                                    bucket.clone(),
-                                    object.clone(),
-                                    self.scan_mode,
-                                    HealChannelPriority::High,
-                                );
-                                (self.update_current_path)(&object).await;
-                                let admission = self
-                                    .send_required_scanner_heal_request(
+                                    self.send_required_scanner_heal_request(
                                         PendingScannerHealKind::Object,
                                         bucket.clone(),
-                                        Some(object.clone()),
+                                        Some(entry.name.clone()),
                                         None,
-                                        request,
+                                        build_object_heal_request(
+                                            bucket.clone(),
+                                            entry.name.clone(),
+                                            None,
+                                            self.scan_mode,
+                                            HealChannelPriority::High,
+                                        ),
                                     )
                                     .await?;
-                                if admission.is_admitted() {
-                                    counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
+                                    found_objects = true;
+                                    continue;
                                 }
+                            };
+
+                            for fiv in fivs.versions {
+                                let version_id = fiv.version_id.and_then(|v| if v.is_nil() { None } else { Some(v.to_string()) });
+                                self.send_required_scanner_heal_request(
+                                    PendingScannerHealKind::Object,
+                                    bucket.clone(),
+                                    Some(entry.name.clone()),
+                                    version_id.clone(),
+                                    build_object_heal_request(
+                                        bucket.clone(),
+                                        entry.name.clone(),
+                                        version_id,
+                                        self.scan_mode,
+                                        HealChannelPriority::High,
+                                    ),
+                                )
+                                .await?;
                                 found_objects = true;
                             }
 

crates/scanner/src/scanner_folder/item_actions.rs

@@ -13,8 +13,6 @@
 // limitations under the License.
 /// Per-object scan actions: ScannerItem, the get-size failure policy, and the heal/ILM admission helpers.
 use super::*;
-#[cfg(test)]
-use rustfs_filemeta::MetadataResolutionParams;
 
 /// Cached folder information for scanning
 #[derive(Clone, Debug)]
@@ -90,22 +88,6 @@ pub(super) fn build_object_heal_request(
     }
 }
 
-/// Build the versionless inspection request used when discovery cannot prove
-/// a destructive version identity (for example an unversioned object or a
-/// bounded candidate overflow). The explicit flag is the fail-closed safety
-/// boundary; callers must not reconstruct it with the destructive default.
-pub(super) fn build_non_destructive_object_heal_request(
-    bucket: String,
-    object: String,
-    scan_mode: HealScanMode,
-    priority: HealChannelPriority,
-) -> HealChannelRequest {
-    let mut request = build_object_heal_request(bucket, object, None, scan_mode, priority);
-    request.remove_corrupted = Some(false);
-    request
-}
-
-#[cfg(test)]
 pub(super) fn resolve_object_heal_entry(
     entries: &MetaCacheEntries,
     resolver: MetadataResolutionParams,

crates/scanner/src/scanner_folder/ledger.rs

@@ -105,29 +105,6 @@ impl FolderScanner {
         }
     }
 
-    /// Preserve the discovery reason when a candidate could not be admitted
-    /// immediately. The existing string field is intentionally reused so the
-    /// scanner's map-encoded cache schema stays backward compatible.
-    pub(super) fn mark_pending_scanner_heal_reason(
-        &mut self,
-        kind: PendingScannerHealKind,
-        bucket: &str,
-        object: Option<&str>,
-        version_id: Option<&str>,
-        reason: &str,
-    ) {
-        if let Some(entry) = self
-            .new_cache
-            .info
-            .pending_heals
-            .iter_mut()
-            .find(|entry| pending_scanner_heal_matches(entry, kind, bucket, object, version_id))
-        {
-            entry.last_admission_reason = reason.to_string();
-            self.sync_pending_heals();
-        }
-    }
-
     pub(super) fn prune_pending_scanner_heals(&mut self) {
         let now = Self::now_secs();
         let before_expiry = self.new_cache.info.pending_heals.len();
@@ -328,22 +305,13 @@ pub(super) fn build_pending_scanner_heal_request(entry: &PendingScannerHeal) ->
     match entry.kind {
         PendingScannerHealKind::Bucket => Some(build_bucket_heal_request(entry.bucket.clone(), HealChannelPriority::High)),
         PendingScannerHealKind::Object => entry.object.as_ref().map(|object| {
-            if entry.version_id.is_none() {
-                build_non_destructive_object_heal_request(
-                    entry.bucket.clone(),
-                    object.clone(),
-                    entry.scan_mode,
-                    HealChannelPriority::High,
-                )
-            } else {
-                build_object_heal_request(
-                    entry.bucket.clone(),
-                    object.clone(),
-                    entry.version_id.clone(),
-                    entry.scan_mode,
-                    HealChannelPriority::High,
-                )
-            }
+            build_object_heal_request(
+                entry.bucket.clone(),
+                object.clone(),
+                entry.version_id.clone(),
+                entry.scan_mode,
+                HealChannelPriority::High,
+            )
         }),
     }
 }

crates/scanner/src/scanner_folder/tests.rs

@@ -17,7 +17,7 @@ use crate::SCANNER_SLEEPER;
 use super::*;
 use crate::storage_api::VersionPurgeStatusType;
 use crate::{DiskOption, Endpoint, STORAGE_FORMAT_FILE, TierStats, new_disk, storageclass};
-use rustfs_filemeta::{FileInfo, FileMeta, MetadataResolutionParams};
+use rustfs_filemeta::{FileInfo, FileMeta};
 use std::io::Write;
 #[cfg(unix)]
 use std::os::unix::fs::{PermissionsExt, symlink};
@@ -982,21 +982,6 @@ fn test_build_object_heal_request_omits_nil_version_id() {
     assert_eq!(request.recreate_missing, Some(false));
 }
 
-#[test]
-fn test_build_non_destructive_object_heal_request_disables_removal() {
-    let request = build_non_destructive_object_heal_request(
-        "bucket".to_string(),
-        "path/to/object".to_string(),
-        HealScanMode::Deep,
-        HealChannelPriority::High,
-    );
-
-    assert_eq!(request.object_version_id, None);
-    assert_eq!(request.remove_corrupted, Some(false));
-    assert_eq!(request.recreate_missing, Some(false));
-    assert_eq!(request.source, HealRequestSource::Scanner);
-}
-
 #[test]
 fn test_build_bucket_heal_request_disables_recreate_for_scanner() {
     let request = build_bucket_heal_request("bucket".to_string(), HealChannelPriority::Low);
@@ -1136,42 +1121,6 @@ fn test_pending_heal_reconstructs_object_request_with_version() {
     assert_eq!(request.source, HealRequestSource::Scanner);
 }
 
-#[test]
-fn test_pending_heal_reconstructs_unversioned_request_without_removal() {
-    let pending = pending_heal(PendingScannerHealKind::Object, "bucket", Some("object"), None, 1, 1);
-
-    let request = build_pending_scanner_heal_request(&pending).expect("unversioned object request should rebuild");
-
-    assert!(request.object_version_id.is_none());
-    assert_eq!(request.remove_corrupted, Some(false));
-    assert_eq!(request.recreate_missing, Some(false));
-}
-
-#[tokio::test]
-async fn test_pending_heal_reason_preserves_sub_quorum_discovery() {
-    let (mut scanner, temp_dir) = build_test_scanner().await;
-    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
-
-    scanner.update_pending_scanner_heal_after_admission(
-        PendingScannerHealKind::Object,
-        "bucket",
-        Some("object"),
-        Some("version-a"),
-        HealScanMode::Deep,
-        HealAdmissionResult::Full,
-    );
-    scanner.mark_pending_scanner_heal_reason(
-        PendingScannerHealKind::Object,
-        "bucket",
-        Some("object"),
-        Some("version-a"),
-        "sub_quorum_metadata",
-    );
-
-    assert_eq!(scanner.new_cache.info.pending_heals.len(), 1);
-    assert_eq!(scanner.new_cache.info.pending_heals[0].last_admission_reason, "sub_quorum_metadata");
-}
-
 #[test]
 fn test_pending_heal_retry_candidates_respect_cap_and_order() {
     let pending: Vec<PendingScannerHeal> = (0..(MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET + 2))
@@ -1374,20 +1323,6 @@ fn metadata_for_object(bucket: &str, object: &str) -> Vec<u8> {
     meta.marshal_msg().expect("test metadata should marshal")
 }
 
-fn metadata_for_object_version(bucket: &str, object: &str, version_id: Option<Uuid>) -> Vec<u8> {
-    let mut file_info = FileInfo::new(object, 4, 2);
-    file_info.volume = bucket.to_string();
-    file_info.name = object.to_string();
-    file_info.version_id = version_id;
-    file_info.versioned = version_id.is_some();
-    file_info.mod_time = Some(OffsetDateTime::now_utc());
-    file_info.size = 1;
-
-    let mut meta = FileMeta::new();
-    meta.add_version(file_info).expect("test metadata version should be accepted");
-    meta.marshal_msg().expect("test metadata should marshal")
-}
-
 async fn write_test_object_metadata(root: &std::path::Path, bucket: &str, object: &str) {
     write_test_object_metadata_bytes(root, bucket, object, &metadata_for_object(bucket, object)).await;
 }
@@ -1791,21 +1726,12 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
     let _guard = TestGuard::new(60, 100, &mut scanner, temp_dir.clone());
     let heal_starts = Arc::new(AtomicUsize::new(0));
     let heal_starts_clone = heal_starts.clone();
-    let healed_versions = Arc::new(Mutex::new(Vec::<Option<String>>::new()));
-    let healed_versions_clone = healed_versions.clone();
     let mut heal_rx =
         rustfs_common::heal_channel::init_heal_channel().expect("heal channel should initialize once for scanner tests");
     let _heal_responder = tokio::spawn(async move {
         while let Some(command) = heal_rx.recv().await {
-            if let rustfs_common::heal_channel::HealChannelCommand::Start {
-                request, response_tx, ..
-            } = command
-            {
+            if let rustfs_common::heal_channel::HealChannelCommand::Start { response_tx, .. } = command {
                 heal_starts_clone.fetch_add(1, Ordering::Relaxed);
-                healed_versions_clone
-                    .lock()
-                    .expect("heal version capture lock should not be poisoned")
-                    .push(request.object_version_id);
                 let _ = response_tx.send(Ok(HealAdmissionResult::Accepted));
             }
         }
@@ -1813,18 +1739,13 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
 
     let bucket = "src-archive";
     let object = "snapshots/37b3f20d941e2f5e6d99114d9bb2f3e67a8a2e5c9c4c5a1b0d6e7f8091a2b3c4";
-    let orphan_version = Uuid::from_u128(0x1934);
-    let shared_version = Uuid::from_u128(0x1935);
-    let orphan_metadata = metadata_for_object_version(bucket, object, Some(orphan_version));
-    let shared_metadata = metadata_for_object_version(bucket, object, Some(shared_version));
-    write_test_object_metadata_bytes(&temp_dir, bucket, object, &orphan_metadata).await;
-    let mut expected_metadata = vec![(temp_dir.join(bucket).join(object).join("xl.meta"), orphan_metadata.clone())];
+    let metadata = metadata_for_object(bucket, object);
+    write_test_object_metadata_bytes(&temp_dir, bucket, object, &metadata).await;
 
     let mut disks = vec![scanner.local_disk.clone()];
     for disk_name in ["disk2", "disk3", "disk4"] {
         let disk_root = temp_dir.join(disk_name);
-        write_test_object_metadata_bytes(&disk_root, bucket, object, &shared_metadata).await;
-        expected_metadata.push((disk_root.join(bucket).join(object).join("xl.meta"), shared_metadata.clone()));
+        write_test_object_metadata_bytes(&disk_root, bucket, object, &metadata).await;
         let endpoint = Endpoint::try_from(disk_root.to_string_lossy().as_ref()).expect("failed to create extra disk endpoint");
         let disk = new_disk(
             &endpoint,
@@ -1873,29 +1794,8 @@ async fn test_scan_folder_exits_when_abandoned_child_listing_finishes() {
         .new_cache
         .checked_flatten(bucket)
         .expect("healed cache must contain canonical child links");
-    // The fixture intentionally exposes two divergent version histories, so
-    // the scanner keeps both logical versions visible while discovering heals.
-    assert_eq!(root.objects, 2);
+    assert_eq!(root.objects, 1);
     assert!(heal_starts.load(Ordering::Relaxed) > 0, "test must execute the heal child-link path");
-    let orphan_version_text = orphan_version.to_string();
-    assert!(
-        healed_versions
-            .lock()
-            .expect("heal version capture lock should not be poisoned")
-            .iter()
-            .any(|version| version.as_deref() == Some(orphan_version_text.as_str())),
-        "sub-quorum orphan version must be submitted as an exact heal candidate"
-    );
-    for (path, expected) in expected_metadata {
-        assert_eq!(
-            tokio::fs::read(&path)
-                .await
-                .expect("scanner discovery must not delete metadata"),
-            expected,
-            "scanner discovery must not modify candidate metadata: {}",
-            path.display()
-        );
-    }
 }
 
 #[tokio::test]

rustfs/src/admin/handlers/mod.rs

@@ -126,6 +126,7 @@ mod tests {
         let _list_remote_target_handler = replication::ListRemoteTargetHandler {};
         let _remove_remote_target_handler = replication::RemoveRemoteTargetHandler {};
         let _scanner_status_handler = scanner::ScannerStatusHandler {};
+        let _scanner_cycle_state_reset_handler = scanner::ScannerCycleStateResetHandler {};
         let _ilm_expiry_status_handler = scanner::IlmExpiryStatusHandler {};
         let _manual_transition_handler = ilm_transition::ManualTransitionRunHandler {};
         let _manual_transition_status_handler = ilm_transition::ManualTransitionJobStatusHandler {};

rustfs/src/admin/handlers/scanner.rs

@@ -13,8 +13,11 @@
 // limitations under the License.
 
 use crate::admin::auth::authorize_admin_request;
+use crate::admin::handlers::supervise_admin_mutation;
 use crate::admin::router::{AdminOperation, Operation, S3Router};
-use crate::admin::runtime_sources::current_scanner_metrics_report;
+use crate::admin::runtime_sources::{
+    app_context_from_req, current_object_store_handle_for_context, current_scanner_metrics_report,
+};
 use crate::module_switches::{ENV_SCANNER_ENABLED, scanner_enabled_from_env};
 use crate::server::ADMIN_PREFIX;
 use chrono::Utc;
@@ -22,11 +25,13 @@ use http::{HeaderMap, HeaderValue};
 use hyper::{Method, StatusCode};
 use matchit::Params;
 use rustfs_common::metrics::{ScannerLifecycleExpirySnapshot, ScannerMaintenanceControlSnapshot, ScannerMetricsReport};
+use rustfs_config::MAX_ADMIN_REQUEST_BODY_SIZE;
 use rustfs_credentials::Credentials;
 use rustfs_policy::policy::action::{Action, AdminAction};
 use s3s::header::CONTENT_TYPE;
 use s3s::{Body, S3Error, S3ErrorCode, S3Request, S3Response, S3Result, s3_error};
-use serde::Serialize;
+use serde::{Deserialize, Serialize};
+use tokio_util::sync::CancellationToken;
 
 const JSON_CONTENT_TYPE: &str = "application/json";
 
@@ -38,6 +43,13 @@ struct ScannerStatusResponse {
     metrics: ScannerMetricsReport,
     cycle_schedule: rustfs_scanner::ScannerCycleScheduleStatus,
     runtime_config: rustfs_scanner::runtime_config::ScannerRuntimeConfigStatus,
+    cycle_recovery: rustfs_scanner::ScannerCycleRecoveryStatus,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct ScannerCycleResetRequest {
+    mode: String,
 }
 
 #[derive(Debug, Serialize)]
@@ -117,6 +129,7 @@ fn scanner_status_response(
         metrics,
         cycle_schedule,
         runtime_config,
+        cycle_recovery: rustfs_scanner::scanner::scanner_cycle_recovery_status(),
     }
 }
 
@@ -144,6 +157,11 @@ pub fn register_scanner_route(r: &mut S3Router<AdminOperation>) -> std::io::Resu
         format!("{ADMIN_PREFIX}/v3/scanner/status").as_str(),
         AdminOperation(&ScannerStatusHandler {}),
     )?;
+    r.insert(
+        Method::POST,
+        format!("{ADMIN_PREFIX}/v3/scanner/cycle-state/reset").as_str(),
+        AdminOperation(&ScannerCycleStateResetHandler {}),
+    )?;
     r.insert(
         Method::GET,
         format!("{ADMIN_PREFIX}/v3/ilm/expiry/status").as_str(),
@@ -163,6 +181,13 @@ async fn validate_scanner_status_request(req: &S3Request<Body>) -> S3Result<Cred
     authorize_admin_request(req, vec![Action::AdminAction(AdminAction::ServerInfoAdminAction)]).await
 }
 
+async fn validate_scanner_reset_request(req: &S3Request<Body>) -> S3Result<Credentials> {
+    if req.credentials.is_none() {
+        return Err(s3_error!(InvalidRequest, "missing credentials"));
+    }
+    authorize_admin_request(req, vec![Action::AdminAction(AdminAction::ConfigUpdateAdminAction)]).await
+}
+
 fn json_response(body: Vec<u8>) -> S3Result<S3Response<(StatusCode, Body)>> {
     let mut headers = HeaderMap::new();
     let content_type = HeaderValue::from_str(JSON_CONTENT_TYPE)
@@ -192,6 +217,37 @@ impl Operation for ScannerStatusHandler {
 
 pub struct IlmExpiryStatusHandler {}
 
+pub struct ScannerCycleStateResetHandler {}
+
+#[async_trait::async_trait]
+impl Operation for ScannerCycleStateResetHandler {
+    async fn call(&self, mut req: S3Request<Body>, _params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
+        let _cred = validate_scanner_reset_request(&req).await?;
+        let body = req
+            .input
+            .store_all_limited(MAX_ADMIN_REQUEST_BODY_SIZE)
+            .await
+            .map_err(|err| S3Error::with_message(S3ErrorCode::InvalidRequest, format!("invalid reset request body: {err}")))?;
+        let reset = serde_json::from_slice::<ScannerCycleResetRequest>(&body)
+            .map_err(|err| S3Error::with_message(S3ErrorCode::InvalidRequest, format!("invalid reset request body: {err}")))?;
+        if reset.mode != "full-rescan" {
+            return Err(S3Error::with_message(S3ErrorCode::InvalidRequest, "reset mode must be full-rescan"));
+        }
+        let context = app_context_from_req(&req)
+            .ok_or_else(|| S3Error::with_message(S3ErrorCode::InternalError, "storage layer not initialized"))?;
+        let store = current_object_store_handle_for_context(Some(context.as_ref()))
+            .ok_or_else(|| S3Error::with_message(S3ErrorCode::InternalError, "storage layer not initialized"))?;
+        supervise_admin_mutation("scanner cycle state reset", async move {
+            rustfs_scanner::scanner::reset_scanner_cycle_recovery(CancellationToken::new(), store)
+                .await
+                .map_err(|err| S3Error::with_message(S3ErrorCode::InternalError, err.to_string()))?;
+            Ok::<_, S3Error>(())
+        })
+        .await?;
+        json_response(br#"{"status":"reset","mode":"full-rescan"}"#.to_vec())
+    }
+}
+
 #[async_trait::async_trait]
 impl Operation for IlmExpiryStatusHandler {
     async fn call(&self, req: S3Request<Body>, _params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
@@ -237,6 +293,38 @@ mod tests {
         assert_eq!(err.message(), Some("missing credentials"));
     }
 
+    #[tokio::test]
+    async fn scanner_reset_gate_rejects_missing_credentials() {
+        let req = S3Request {
+            input: Body::from(String::new()),
+            method: Method::POST,
+            uri: http::Uri::from_static("/rustfs/admin/v3/scanner/cycle-state/reset"),
+            headers: HeaderMap::new(),
+            extensions: http::Extensions::new(),
+            credentials: None,
+            region: None,
+            service: None,
+            trailing_headers: None,
+        };
+
+        let err = validate_scanner_reset_request(&req)
+            .await
+            .expect_err("a reset request without credentials must be rejected");
+        assert_eq!(err.code(), &S3ErrorCode::InvalidRequest);
+        assert_eq!(err.message(), Some("missing credentials"));
+    }
+
+    #[test]
+    fn admin_reset_requires_full_rescan_or_verified_cursor() {
+        let full_rescan: ScannerCycleResetRequest =
+            serde_json::from_str(r#"{"mode":"full-rescan"}"#).expect("full rescan must be accepted");
+        assert_eq!(full_rescan.mode, "full-rescan");
+        let cursor: ScannerCycleResetRequest =
+            serde_json::from_str(r#"{"mode":"cursor"}"#).expect("mode validation belongs to the handler");
+        assert_ne!(cursor.mode, "full-rescan");
+        assert!(serde_json::from_str::<ScannerCycleResetRequest>(r#"{"mode":"full-rescan","cursor":"untrusted"}"#).is_err());
+    }
+
     #[test]
     fn scanner_disabled_reason_reports_startup_env_key() {
         assert_eq!(scanner_disabled_reason(true), None);
@@ -304,6 +392,11 @@ mod tests {
         assert_eq!(encoded["cycle_schedule"]["effective_interval_seconds"], 0);
         assert_eq!(encoded["cycle_schedule"]["clean_idle_backoff_enabled"], false);
         assert_eq!(encoded["cycle_schedule"]["clean_idle_backoff_multiplier"], 1);
+        assert_eq!(encoded["cycle_recovery"]["state"], "healthy");
+        assert_eq!(
+            encoded["cycle_recovery"]["quarantine_path"],
+            rustfs_scanner::DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()
+        );
     }
 
     #[test]

rustfs/src/admin/route_policy.rs

@@ -428,6 +428,12 @@ pub const ADMIN_ROUTE_POLICY_SPECS: &[AdminRouteSpec] = &[
     admin(HttpMethod::Get, "/rustfs/admin/v3/config", CONFIG_UPDATE, RouteRiskLevel::High),
     admin(HttpMethod::Put, "/rustfs/admin/v3/config", CONFIG_UPDATE, RouteRiskLevel::High),
     admin(HttpMethod::Get, "/rustfs/admin/v3/scanner/status", SERVER_INFO, RouteRiskLevel::Sensitive),
+    admin(
+        HttpMethod::Post,
+        "/rustfs/admin/v3/scanner/cycle-state/reset",
+        CONFIG_UPDATE,
+        RouteRiskLevel::High,
+    ),
     admin(
         HttpMethod::Get,
         "/rustfs/admin/v3/ilm/expiry/status",
@@ -2020,6 +2026,12 @@ mod tests {
         assert_not_action(HttpMethod::Get, "/rustfs/admin/v3/ilm/expiry/status", SET_TIER);
     }
 
+    #[test]
+    fn route_policy_requires_config_update_for_scanner_cycle_reset() {
+        assert_action(HttpMethod::Post, "/rustfs/admin/v3/scanner/cycle-state/reset", CONFIG_UPDATE);
+        assert_not_action(HttpMethod::Post, "/rustfs/admin/v3/scanner/cycle-state/reset", SERVER_INFO);
+    }
+
     #[test]
     fn route_policy_uses_tier_actions_for_transition_routes() {
         assert_action(HttpMethod::Post, "/rustfs/admin/v3/ilm/transition/run", SET_TIER);

rustfs/src/admin/route_registration_test.rs

@@ -243,6 +243,7 @@ fn expected_admin_route_matrix() -> Vec<RouteMatrixEntry> {
         admin_route(Method::GET, "/v3/config"),
         admin_route(Method::PUT, "/v3/config"),
         admin_route(Method::GET, "/v3/scanner/status"),
+        admin_route(Method::POST, "/v3/scanner/cycle-state/reset"),
         admin_route(Method::GET, "/v3/audit/target/list"),
         admin_route_sample(
             Method::PUT,
@@ -879,6 +880,7 @@ fn test_register_routes_cover_representative_admin_paths() {
     assert_route(&router, Method::GET, &admin_path("/v3/config"));
     assert_route(&router, Method::PUT, &admin_path("/v3/config"));
     assert_route(&router, Method::GET, &admin_path("/v3/scanner/status"));
+    assert_route(&router, Method::POST, &admin_path("/v3/scanner/cycle-state/reset"));
     assert_route(&router, Method::GET, &admin_path("/v3/ilm/expiry/status"));
     assert_route(&router, Method::POST, &admin_path("/v3/ilm/transition/run"));
     assert_route(
@@ -1367,6 +1369,7 @@ fn test_admin_alias_paths_match_existing_admin_routes() {
         (Method::GET, compat_admin_alias_path("/v3/config")),
         (Method::PUT, compat_admin_alias_path("/v3/config")),
         (Method::GET, compat_admin_alias_path("/v3/scanner/status")),
+        (Method::POST, compat_admin_alias_path("/v3/scanner/cycle-state/reset")),
         (Method::GET, compat_admin_alias_path("/v3/ilm/expiry/status")),
     ] {
         assert!(

scripts/test/vault_ha_kms_live.sh

@@ -241,7 +241,7 @@ env \
   RUSTFS_TEST_VAULT_FAILOVER_MARKER="$MARKER" \
   RUSTFS_TEST_VAULT_OLD_LEADER="$OLD_LEADER" \
   cargo test -p rustfs-kms --test vault_ha_failover_live \
-    vault_raft_leader_failure_preserves_kv2_and_transit_decrypts -- \
+    vault_raft_leader_failure_recovers_kv2_and_transit_decrypts -- \
     --ignored --nocapture --test-threads=1 &
 TEST_PID=$!