fix(heal): require current checkpoint digest

The inline_block threshold used floor division (DEFAULT_INLINE_OBJECT_BUDGET
/ data_shards) while shard_file_size uses ceiling division (div_ceil). For
EC 12:4 with 256KiB objects, this caused a 1-byte discrepancy:
- inline_block = 262144 / 12 = 21845 (floor)
- shard_file_size = 262144.div_ceil(12) = 21846 (ceil)
- should_inline(21846, 12, false) = false (wrong!)

Fix by using div_ceil for the inline_block calculation, so both sides
use the same rounding and the inline path is correctly triggered.

Co-authored-by: heihutu <heihutu@gmail.com>

Cargo.lock

@@ -9587,6 +9587,7 @@ dependencies = [
  "serde",
  "serde_json",
  "serial_test",
+ "sha2 0.11.0",
  "temp-env",
  "tempfile",
  "thiserror 2.0.20",

crates/data-usage/src/data_usage.rs

@@ -585,9 +585,12 @@ impl VersionsHistogram {
     }
 }
 
-/// Replication statistics for a single target
-#[derive(Debug, Default, Clone, Serialize, Deserialize)]
-pub struct ReplicationStats {
+/// Replication statistics for a single target.
+///
+/// Renamed from `ReplicationStats`; serde field names are preserved
+/// byte-identically to maintain wire compatibility with existing snapshots.
+#[derive(Debug, Default, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct ReplicationTargetUsage {
     pub pending_size: u64,
     pub replicated_size: u64,
     pub failed_size: u64,
@@ -600,7 +603,7 @@ pub struct ReplicationStats {
     pub replicated_count: u64,
 }
 
-impl ReplicationStats {
+impl ReplicationTargetUsage {
     pub fn is_empty(&self) -> bool {
         let Self {
             pending_size,
@@ -636,7 +639,7 @@ impl ReplicationStats {
 /// Replication statistics for all targets
 #[derive(Debug, Default, Clone, Serialize, Deserialize)]
 pub struct ReplicationAllStats {
-    pub targets: HashMap<String, ReplicationStats>,
+    pub targets: HashMap<String, ReplicationTargetUsage>,
     pub replica_size: u64,
     pub replica_count: u64,
 }
@@ -649,7 +652,7 @@ impl ReplicationAllStats {
             targets,
         } = self;
 
-        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationStats::is_empty)
+        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationTargetUsage::is_empty)
     }
 
     #[deprecated(note = "use is_empty instead")]
@@ -2466,7 +2469,7 @@ mod tests {
 
     #[test]
     fn replication_stats_empty_checks_every_field() {
-        type SetField = fn(&mut ReplicationStats);
+        type SetField = fn(&mut ReplicationTargetUsage);
 
         let cases: [(&str, SetField); 10] = [
             ("pending_size", |stats| stats.pending_size = 1),
@@ -2481,9 +2484,9 @@ mod tests {
             ("replicated_count", |stats| stats.replicated_count = 1),
         ];
 
-        assert!(ReplicationStats::default().is_empty());
+        assert!(ReplicationTargetUsage::default().is_empty());
         for (field, set_nonzero) in cases {
-            let mut stats = ReplicationStats::default();
+            let mut stats = ReplicationTargetUsage::default();
             set_nonzero(&mut stats);
             assert!(!stats.is_empty(), "{field} must make replication stats non-empty");
         }
@@ -2514,17 +2517,17 @@ mod tests {
         }
 
         let empty_targets = ReplicationAllStats {
-            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationStats::default())]),
+            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationTargetUsage::default())]),
             ..Default::default()
         };
         assert!(empty_targets.is_empty(), "all-empty targets must keep aggregate stats empty");
 
         let stats = ReplicationAllStats {
             targets: HashMap::from([
-                ("arn:test:empty".to_string(), ReplicationStats::default()),
+                ("arn:test:empty".to_string(), ReplicationTargetUsage::default()),
                 (
                     "arn:test:non-empty".to_string(),
-                    ReplicationStats {
+                    ReplicationTargetUsage {
                         pending_count: 1,
                         ..Default::default()
                     },
@@ -2565,7 +2568,7 @@ mod tests {
                 replication_stats: Some(ReplicationAllStats {
                     targets: HashMap::from([(
                         "arn:test:pending".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                             pending_count: 1,
                             ..Default::default()
                         },
@@ -2714,7 +2717,7 @@ mod tests {
                 targets: HashMap::from([
                     (
                         "arn:self-only".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                             pending_size: 7,
                             pending_count: 1,
                             ..Default::default()
@@ -2722,7 +2725,7 @@ mod tests {
                     ),
                     (
                         "arn:shared".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                             failed_size: 3,
                             failed_count: 1,
                             missed_threshold_size: 2,
@@ -2741,7 +2744,7 @@ mod tests {
                 targets: HashMap::from([
                     (
                         "arn:shared".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                             failed_size: 5,
                             failed_count: 2,
                             after_threshold_size: 4,
@@ -2751,7 +2754,7 @@ mod tests {
                     ),
                     (
                         "arn:other-only".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                             replicated_size: 11,
                             replicated_count: 3,
                             ..Default::default()
@@ -2993,7 +2996,9 @@ mod tests {
     fn replication_target_deserialization_preserves_large_historical_maps() {
         let mut stats = ReplicationAllStats::default();
         for index in 0..=1024 {
-            stats.targets.insert(format!("target-{index}"), ReplicationStats::default());
+            stats
+                .targets
+                .insert(format!("target-{index}"), ReplicationTargetUsage::default());
         }
         let encoded = rmp_serde::to_vec_named(&stats).expect("large replication target fixture should encode");
         let decoded = rmp_serde::from_slice::<ReplicationAllStats>(&encoded)
@@ -3002,6 +3007,47 @@ mod tests {
         assert_eq!(decoded.targets.len(), stats.targets.len());
     }
 
+    /// Round-trip test: encoding a [`ReplicationTargetUsage`] and decoding it back
+    /// must produce the exact same value.  This guards against accidental serde
+    /// field-name drift during the `ReplicationStats` -> `ReplicationTargetUsage`
+    /// rename.  Wire-level field names are the serialized Rust field identifiers,
+    /// which must remain byte-identical.
+    #[test]
+    fn replication_target_usage_rmp_round_trip() {
+        let original = ReplicationTargetUsage {
+            pending_size: 100,
+            replicated_size: 2_000,
+            failed_size: 50,
+            failed_count: 3,
+            pending_count: 7,
+            missed_threshold_size: 11,
+            after_threshold_size: 22,
+            missed_threshold_count: 1,
+            after_threshold_count: 2,
+            replicated_count: 99,
+        };
+
+        let buf = rmp_serde::to_vec_named(&original).expect("encode ReplicationTargetUsage to msgpack");
+        let decoded: ReplicationTargetUsage = rmp_serde::from_slice(&buf).expect("decode ReplicationTargetUsage from msgpack");
+        assert_eq!(original, decoded, "round-trip through rmp must preserve every field");
+
+        // Also verify that encoding as an unnamed sequence and then decoding
+        // with named fields produces the correct mapping (this catches reordering).
+        let named_buf = rmp_serde::to_vec_named(&original).expect("re-encode for field-name pinning");
+        // Spot-check that known field names appear in the named encoding.
+        let named_str = String::from_utf8_lossy(&named_buf);
+        assert!(named_str.contains("pending_size"), "field 'pending_size' must survive the rename");
+        assert!(named_str.contains("replicated_size"), "field 'replicated_size' must survive the rename");
+        assert!(
+            named_str.contains("missed_threshold_size"),
+            "field 'missed_threshold_size' must survive the rename"
+        );
+        assert!(
+            named_str.contains("after_threshold_count"),
+            "field 'after_threshold_count' must survive the rename"
+        );
+    }
+
     #[test]
     fn checked_merge_rejects_noncanonical_histograms_without_mutation() {
         let mut entry = DataUsageEntry {

crates/e2e_test/src/heal_erasure_disk_rebuild_test.rs

@@ -380,10 +380,24 @@ mod tests {
         cluster.start_node(1).await?;
 
         let status_url = format!("{}/rustfs/admin/v3/background-heal/status", cluster.nodes[0].url);
-        let status_body = signed_admin_post(&status_url, None, &cluster.access_key, &cluster.secret_key).await?;
-        assert!(
-            !status_body.contains("MissingContentLength"),
-            "background heal status should not fail without an explicit Content-Length: {status_body}"
+        let mut recovered = serde_json::Value::Null;
+        for _ in 0..60 {
+            let status_body = signed_admin_post(&status_url, None, &cluster.access_key, &cluster.secret_key).await?;
+            assert!(
+                !status_body.contains("MissingContentLength"),
+                "background heal status should not fail without an explicit Content-Length: {status_body}"
+            );
+            recovered = serde_json::from_str(&status_body)
+                .map_err(|err| format!("background heal status is not JSON ({err}): {status_body}"))?;
+            if recovered["clusterStatusComplete"] == serde_json::Value::Bool(true) {
+                break;
+            }
+            sleep(Duration::from_secs(1)).await;
+        }
+        assert_eq!(
+            recovered["clusterStatusComplete"],
+            serde_json::Value::Bool(true),
+            "cluster heal status should recover before root heal starts: {recovered}"
         );
 
         let heal_body = r#"{"recursive":true,"dryRun":false,"remove":false,"recreate":true,"scanMode":2,"updateParity":false,"nolock":false}"#;

crates/ecstore/src/api/mod.rs

@@ -317,8 +317,6 @@ pub mod config {
 }
 
 pub mod data_usage {
-    #[cfg(feature = "test-util")]
-    pub use crate::data_usage::seed_bucket_usage_memory_for_test;
     pub use crate::data_usage::{
         DATA_USAGE_CACHE_NAME, apply_bucket_usage_memory_overlay, compute_bucket_usage,
         init_compression_total_memory_from_backend, invalidate_admin_data_usage_snapshot_cache,
@@ -330,6 +328,8 @@ pub mod data_usage {
         remove_bucket_usage_from_backend, replace_bucket_usage_memory_from_info, store_compression_total_in_backend,
         store_data_usage_in_backend,
     };
+    #[cfg(feature = "test-util")]
+    pub use crate::data_usage::{get_bucket_usage_memory, seed_bucket_usage_memory_for_test};
 }
 
 pub mod disk {

crates/ecstore/src/bucket/bucket_target_sys.rs

@@ -3425,6 +3425,44 @@ mod tests {
         assert!(mutexes.contains_key("second"));
     }
 
+    #[tokio::test]
+    async fn update_all_targets_publishes_disable_proxy_on_target_client() {
+        // The read-proxy selector (replication_proxy::get_proxy_targets) skips
+        // targets whose TargetClient carries disable_proxy — the persisted
+        // per-target opt-out must survive client publication.
+        let sys = BucketTargetSys::default();
+        let target = |arn: &str, disable_proxy: bool| BucketTarget {
+            arn: arn.to_string(),
+            endpoint: "192.168.1.10:9000".to_string(),
+            target_bucket: "target-bucket".to_string(),
+            region: "us-east-1".to_string(),
+            disable_proxy,
+            credentials: Some(Credentials {
+                access_key: "access".to_string(),
+                secret_key: "secret".to_string(),
+                session_token: None,
+                expiration: None,
+            }),
+            ..Default::default()
+        };
+        let targets = BucketTargets {
+            targets: vec![target("arn:proxied", false), target("arn:opted-out", true)],
+        };
+
+        sys.update_all_targets("bucket", Some(&targets)).await;
+
+        let proxied = sys
+            .get_remote_target_client("bucket", "arn:proxied")
+            .await
+            .expect("client should be published");
+        assert!(!proxied.disable_proxy);
+        let opted_out = sys
+            .get_remote_target_client("bucket", "arn:opted-out")
+            .await
+            .expect("client should be published");
+        assert!(opted_out.disable_proxy, "disable_proxy must reach the published TargetClient");
+    }
+
     #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
     async fn target_updates_serialize_client_build_through_publication_per_bucket() {
         let sys = Arc::new(BucketTargetSys::default());

crates/ecstore/src/bucket/replication/replication_pool.rs

@@ -2855,7 +2855,7 @@ fn replicate_object_info_from_object_info(
         .map(|v| OffsetDateTime::parse(&v, &Rfc3339).unwrap_or(OffsetDateTime::UNIX_EPOCH));
     let mut rstate = oi.replication_state();
     rstate.replicate_decision_str = dsc.to_string();
-    let asz = oi.get_actual_size().unwrap_or_default();
+    let asz = oi.get_actual_size_or_physical();
     let ssec = replication_object_is_ssec_encrypted(&oi.user_defined);
     let checksum = if ssec { oi.checksum.clone() } else { None };
 

crates/ecstore/src/bucket/replication/replication_resyncer.rs

@@ -1412,7 +1412,7 @@ pub async fn get_heal_replicate_object_info(oi: &ObjectInfo, rcfg: &ReplicationC
     };
     let mut replication_state = oi.replication_state();
     replication_state.replicate_decision_str = dsc.to_string();
-    let actual_size = oi.get_actual_size().unwrap_or_default();
+    let actual_size = oi.get_actual_size_or_physical();
 
     Ok(ReplicateObjectInfo {
         name: oi.name.clone(),

crates/ecstore/src/bucket/replication/replication_target_boundary.rs

@@ -389,7 +389,7 @@ fn replication_source_object(object_info: &ObjectInfo) -> ReplicationSourceObjec
             .map(|mod_time| OffsetDateTime::from_unix_timestamp(mod_time.unix_timestamp()).unwrap_or(mod_time)),
         version_id: object_info.version_id.map(|version_id| version_id.to_string()),
         etag: object_info.etag.as_deref(),
-        actual_size: object_info.get_actual_size().unwrap_or_default(),
+        actual_size: object_info.get_actual_size_or_physical(),
         delete_marker: object_info.delete_marker,
         content_type: object_info.content_type.as_deref(),
         content_encoding: object_info.content_encoding.as_deref(),
@@ -542,6 +542,20 @@ mod tests {
         assert!(replication_target_head_is_newer_null_version(&source, &target));
     }
 
+    #[test]
+    fn replication_source_uses_physical_size_for_unknown_compressed_object() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(&mut metadata, rustfs_utils::http::SUFFIX_COMPRESSION, "zstd".to_string());
+        let source = ObjectInfo {
+            size: 128,
+            actual_size: -1,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+
+        assert_eq!(replication_source_object(&source).actual_size, 128);
+    }
+
     #[test]
     fn replication_target_head_content_matches_compare_etag_only() {
         let source = ObjectInfo {

crates/ecstore/src/config/storageclass.rs

@@ -248,10 +248,13 @@ impl Config {
         let shard_size = shard_size as usize;
         // Keep the historical two-data-shard object budget while preventing
         // wider EC layouts from multiplying the maximum inline object size.
+        // Use div_ceil to match the shard_file_size calculation (which also uses
+        // div_ceil), avoiding a 1-byte rounding discrepancy that prevents inline
+        // for objects right at the threshold.
         let inline_block = if self.initialized && self.inline_block_explicit {
             self.inline_block
         } else {
-            (DEFAULT_INLINE_OBJECT_BUDGET / data_shards).min(DEFAULT_INLINE_BLOCK)
+            DEFAULT_INLINE_OBJECT_BUDGET.div_ceil(data_shards).min(DEFAULT_INLINE_BLOCK)
         };
 
         if versioned {

crates/ecstore/src/core/sets.rs

@@ -21,7 +21,7 @@ use crate::storage_api_contracts::{
     bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOptions, MakeBucketOptions},
     list::{StorageListObjectVersionsInfo, StorageListObjectsV2Info, StorageObjectInfoOrErr, StorageWalkOptions},
     multipart::{CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartUploadResult, PartInfo},
-    object::{DeletedObject, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
+    object::{DeleteAccounting, DeletedObject, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
     range::HTTPRangeSpec,
 };
 use crate::{
@@ -414,6 +414,66 @@ fn apply_delete_objects_results(
     }
 }
 
+fn apply_delete_accounting_results(
+    accounting: &mut [Option<DeleteAccounting>],
+    set_objects: &[DelObj],
+    set_accounting: &[Option<DeleteAccounting>],
+) {
+    for (obj, value) in set_objects.iter().zip(set_accounting.iter()) {
+        accounting[obj.orig_idx] = value.clone();
+    }
+}
+
+impl Sets {
+    pub(crate) async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+        let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut del_errs = vec![None; objects.len()];
+        let mut accounting = vec![None; objects.len()];
+        let mut set_obj_map = HashMap::new();
+
+        for (i, obj) in objects.iter().enumerate() {
+            let idx = self.get_hashed_set_index(obj.object_name.as_str());
+            set_obj_map.entry(idx).or_insert_with(Vec::new).push(DelObj {
+                orig_idx: i,
+                obj: obj.clone(),
+            });
+        }
+
+        let max_concurrent = set_obj_map.len().min(num_cpus::get()).max(1);
+        let semaphore = Arc::new(tokio::sync::Semaphore::new(max_concurrent));
+        let mut futures = FuturesUnordered::new();
+        let bucket = bucket.to_owned();
+
+        for (set_index, set_objects) in set_obj_map {
+            let disks = self.get_disks(set_index);
+            let objects = set_objects.iter().map(|entry| entry.obj.clone()).collect::<Vec<_>>();
+            let bucket = bucket.clone();
+            let opts = opts.clone();
+            let semaphore = semaphore.clone();
+            futures.push(async move {
+                let _permit = semaphore
+                    .acquire_owned()
+                    .await
+                    .expect("delete_objects semaphore should remain open");
+                let (deleted, errors, accounting) = disks.delete_objects_with_accounting(&bucket, objects, opts).await;
+                (set_objects, deleted, errors, accounting)
+            });
+        }
+
+        while let Some((set_objects, deleted, errors, set_accounting)) = futures.next().await {
+            apply_delete_objects_results(&mut del_objects, &mut del_errs, &set_objects, &deleted, errors);
+            apply_delete_accounting_results(&mut accounting, &set_objects, &set_accounting);
+        }
+
+        (del_objects, del_errs, accounting)
+    }
+}
+
 #[async_trait::async_trait]
 impl crate::storage_api_contracts::object::ObjectIO for Sets {
     type Error = Error;
@@ -655,65 +715,8 @@ impl crate::storage_api_contracts::object::ObjectOperations for Sets {
         objects: Vec<ObjectToDelete>,
         opts: ObjectOptions,
     ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
-        // Default return value
-        let mut del_objects = vec![DeletedObject::default(); objects.len()];
-
-        let mut del_errs = Vec::with_capacity(objects.len());
-        for _ in 0..objects.len() {
-            del_errs.push(None)
-        }
-
-        let mut set_obj_map = HashMap::new();
-
-        // hash key
-        for (i, obj) in objects.iter().enumerate() {
-            let idx = self.get_hashed_set_index(obj.object_name.as_str());
-
-            if !set_obj_map.contains_key(&idx) {
-                set_obj_map.insert(
-                    idx,
-                    vec![DelObj {
-                        // set_idx: idx,
-                        orig_idx: i,
-                        obj: obj.clone(),
-                    }],
-                );
-            } else if let Some(val) = set_obj_map.get_mut(&idx) {
-                val.push(DelObj {
-                    // set_idx: idx,
-                    orig_idx: i,
-                    obj: obj.clone(),
-                });
-            }
-        }
-
-        let max_concurrent = set_obj_map.len().min(num_cpus::get()).max(1);
-        let semaphore = Arc::new(tokio::sync::Semaphore::new(max_concurrent));
-        let mut futures = FuturesUnordered::new();
-        let bucket = bucket.to_string();
-
-        for (k, v) in set_obj_map {
-            let disks = self.get_disks(k);
-            let objs: Vec<ObjectToDelete> = v.iter().map(|v| v.obj.clone()).collect();
-            let bucket = bucket.clone();
-            let opts = opts.clone();
-            let semaphore = semaphore.clone();
-
-            futures.push(async move {
-                let _permit = semaphore
-                    .acquire_owned()
-                    .await
-                    .expect("delete_objects semaphore should remain open");
-                let (dobjects, errs) = disks.delete_objects(&bucket, objs, opts).await;
-                (v, dobjects, errs)
-            });
-        }
-
-        while let Some((v, dobjects, errs)) = futures.next().await {
-            apply_delete_objects_results(&mut del_objects, &mut del_errs, &v, &dobjects, errs);
-        }
-
-        (del_objects, del_errs)
+        let (deleted, errors, _) = self.delete_objects_with_accounting(bucket, objects, opts).await;
+        (deleted, errors)
     }
 
     #[tracing::instrument(skip(self))]

crates/ecstore/src/data_usage/mod.rs

@@ -1391,7 +1391,37 @@ impl BucketUsageAccumulator {
 }
 
 pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
-    let logical_size = u64::try_from(object.get_actual_size().map_err(Error::other)?).map_err(|_| Error::PartMissingOrCorrupt)?;
+    // A compressed object may carry -1 while the transformed size is unknown
+    // (legacy streaming sentinel). In that case the persisted physical size
+    // is still a valid accounting floor; every other negative value is corrupt.
+    // An explicit negative `actual-size` metadata value is corrupt, however:
+    // the sentinel is only valid in the in-memory/object-part field written by
+    // the legacy streaming path, not as a persisted declared size.
+    let compressed = object.is_compressed();
+    if object.actual_size < -1 || (object.actual_size == -1 && !compressed) {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    if object
+        .parts
+        .iter()
+        .any(|part| part.actual_size < -1 || (part.actual_size < 0 && !compressed))
+    {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    let declared_actual_size = rustfs_utils::http::get_str(&object.user_defined, rustfs_utils::http::SUFFIX_ACTUAL_SIZE)
+        .filter(|value| !value.is_empty());
+    if declared_actual_size
+        .as_deref()
+        .and_then(|value| value.parse::<i64>().ok())
+        .is_some_and(|size| size < 0)
+    {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    let logical_size = match object.get_actual_size().map_err(Error::other)? {
+        size if size == -1 && compressed && declared_actual_size.is_none() => None,
+        size if size >= 0 => Some(u64::try_from(size).map_err(|_| Error::PartMissingOrCorrupt)?),
+        _ => return Err(Error::PartMissingOrCorrupt),
+    };
     let persisted_part_size = if object.parts.is_empty() {
         u64::try_from(object.size).map_err(|_| Error::PartMissingOrCorrupt)?
     } else {
@@ -1399,12 +1429,8 @@ pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
             // Compressed streaming objects persist -1 when the transformed
             // part size is unknown. The physical part size remains a valid
             // quota floor; reject only non-negative values that overflow.
-            let actual_size = if part.actual_size < 0 {
-                if object.is_compressed() {
-                    0
-                } else {
-                    return Err(Error::PartMissingOrCorrupt);
-                }
+            let actual_size = if part.actual_size == -1 {
+                0
             } else {
                 u64::try_from(part.actual_size).map_err(|_| Error::PartMissingOrCorrupt)?
             };
@@ -1412,7 +1438,7 @@ pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
             total.checked_add(part_size).ok_or(Error::PartMissingOrCorrupt)
         })?
     };
-    Ok(logical_size.max(persisted_part_size))
+    Ok(logical_size.unwrap_or(0).max(persisted_part_size))
 }
 
 type UsageVersionPage = StorageListObjectVersionsInfo<ObjectInfo>;
@@ -3320,6 +3346,80 @@ mod tests {
         );
     }
 
+    #[test]
+    fn quota_object_size_accepts_compressed_unknown_actual_size_sentinel() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let object = ObjectInfo {
+            size: 400,
+            actual_size: -1,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+
+        assert_eq!(quota_object_size(&object).expect("compressed sentinel is valid"), 400);
+    }
+
+    #[test]
+    fn quota_object_size_rejects_compressed_part_sum_overflow() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let object = ObjectInfo {
+            size: 1,
+            user_defined: Arc::new(metadata),
+            parts: Arc::new(vec![
+                rustfs_filemeta::ObjectPartInfo {
+                    actual_size: i64::MAX,
+                    ..Default::default()
+                },
+                rustfs_filemeta::ObjectPartInfo {
+                    actual_size: 1,
+                    ..Default::default()
+                },
+            ]),
+            ..Default::default()
+        };
+
+        assert!(matches!(quota_object_size(&object), Err(Error::Io(_))));
+    }
+
+    #[test]
+    fn quota_object_size_rejects_negative_values_other_than_the_compressed_sentinel() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let corrupt_object = ObjectInfo {
+            size: 400,
+            actual_size: -2,
+            user_defined: Arc::new(metadata.clone()),
+            ..Default::default()
+        };
+        assert!(matches!(quota_object_size(&corrupt_object), Err(Error::PartMissingOrCorrupt)));
+
+        let corrupt_part = ObjectInfo {
+            size: 400,
+            user_defined: Arc::new(metadata),
+            parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+                size: 400,
+                actual_size: -2,
+                ..Default::default()
+            }]),
+            ..Default::default()
+        };
+        assert!(matches!(quota_object_size(&corrupt_part), Err(Error::PartMissingOrCorrupt)));
+    }
+
     #[tokio::test]
     #[serial]
     async fn live_bucket_usage_refreshes_are_coalesced_only_while_in_flight() {

crates/ecstore/src/disk/local.rs

@@ -7949,10 +7949,15 @@ impl DiskAPI for LocalDisk {
             use std::io::Write as _;
 
             let file_path = self.io_get_object_path(volume, path)?;
-            let lock_path = file_path.with_extension("rustfs-cas.lock");
             let path = path.to_string();
             let sync_metadata = effective_durability(volume).syncs_commit_metadata();
             return Ok(tokio::task::spawn_blocking(move || {
+                // A persistent directory lock bounds metadata growth. Removing
+                // per-target lock files can split flock ownership across inodes.
+                let lock_path = file_path
+                    .parent()
+                    .ok_or_else(|| std::io::Error::new(ErrorKind::InvalidInput, "conditional file has no parent"))?
+                    .join(".rustfs-cas.lock");
                 let lock = std::fs::OpenOptions::new()
                     .create(true)
                     .truncate(false)
@@ -21841,7 +21846,10 @@ mod test {
         let marker_path = disk
             .get_object_path(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
             .expect("marker path should resolve");
-        let lock_path = marker_path.with_extension("rustfs-cas.lock");
+        let lock_path = marker_path
+            .parent()
+            .expect("marker path should have a parent")
+            .join(".rustfs-cas.lock");
         let lock = std::fs::OpenOptions::new()
             .create(true)
             .truncate(false)

crates/ecstore/src/object_api/types.rs

@@ -689,6 +689,9 @@ impl ObjectInfo {
     }
 
     pub fn get_actual_size(&self) -> std::io::Result<i64> {
+        if self.actual_size < -1 || (self.actual_size == -1 && !self.is_compressed()) {
+            return Err(std::io::Error::other("invalid negative actual size"));
+        }
         if self.actual_size > 0 {
             return Ok(self.actual_size);
         }
@@ -700,10 +703,25 @@ impl ObjectInfo {
                 let size = size_str.parse::<i64>().map_err(|e| std::io::Error::other(e.to_string()))?;
                 return Ok(size);
             }
-            let mut actual_size = 0;
-            self.parts.iter().for_each(|part| {
-                actual_size += part.actual_size;
-            });
+            if self.actual_size == -1 && self.parts.is_empty() {
+                return Ok(-1);
+            }
+            let mut actual_size = 0_i64;
+            let mut unknown = false;
+            for part in self.parts.iter() {
+                match part.actual_size {
+                    -1 => unknown = true,
+                    size if size >= 0 => {
+                        actual_size = actual_size
+                            .checked_add(size)
+                            .ok_or_else(|| std::io::Error::other("compressed actual size overflow"))?;
+                    }
+                    _ => return Err(std::io::Error::other("invalid negative compressed part size")),
+                }
+            }
+            if unknown {
+                return Ok(-1);
+            }
             if actual_size == 0 && actual_size != self.size {
                 return Err(std::io::Error::other(format!("invalid decompressed size {} {}", actual_size, self.size)));
             }
@@ -718,6 +736,18 @@ impl ObjectInfo {
         Ok(self.size)
     }
 
+    /// Returns a non-negative size for client and replication boundaries.
+    ///
+    /// Compressed legacy metadata can retain the internal `-1` unknown-size
+    /// sentinel. Those boundaries cannot emit a negative length, so they use
+    /// the persisted physical size while quota accounting keeps the sentinel
+    /// distinction in [`crate::data_usage::quota_object_size`].
+    pub fn get_actual_size_or_physical(&self) -> i64 {
+        self.get_actual_size()
+            .map(|size| if size >= 0 { size } else { self.size.max(0) })
+            .unwrap_or_else(|_| self.size.max(0))
+    }
+
     pub fn from_file_info(fi: &FileInfo, bucket: &str, object: &str, versioned: bool) -> ObjectInfo {
         let mut version_id = fi.version_id;
 

crates/ecstore/src/set_disk/mod.rs

@@ -97,7 +97,7 @@ use crate::storage_api_contracts::{
         CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartOperations as _, MultipartUploadResult, PartInfo,
     },
     namespace::NamespaceLocking as _,
-    object::{DeletedObject, HTTPPreconditions, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
+    object::{DeleteAccounting, DeletedObject, HTTPPreconditions, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
     range::HTTPRangeSpec,
 };
 use crate::store::utils::is_reserved_or_invalid_bucket;

crates/ecstore/src/set_disk/ops/object.rs

@@ -45,6 +45,7 @@ use crate::bucket::replication::{
     DeleteReplicationConfigSnapshot, ReplicationLifecycleBridge, ReplicationStatusType, VersionPurgeStatusType,
     replication_state_to_filemeta, replication_status_from_filemeta, version_purge_status_to_filemeta,
 };
+use crate::data_usage::quota_object_size;
 use crate::diagnostics::get::GetObjectFailureReason;
 use crate::disk::{DataDirDeleteStatus, OldCurrentSize};
 use crate::error::is_err_invalid_upload_id;
@@ -2122,14 +2123,27 @@ impl SetDisks {
             let erasure = Arc::new(erasure_from_file_info(&fi, false)?);
 
             let put_object_size = known_put_object_storage_size(data.size());
+            let shard_file_size_raw = erasure.shard_file_size(put_object_size);
             let is_inline_buffer =
-                storage_class_config.should_inline(erasure.shard_file_size(put_object_size), erasure.data_shards, opts.versioned);
+                storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);
 
             let collect_stage_timing = rustfs_io_metrics::put_stage_metrics_enabled() || issue3031_diag_enabled();
-            let shard_file_size = erasure.shard_file_size(put_object_size);
+            let shard_file_size = shard_file_size_raw;
             let shard_size = erasure.shard_size();
             let write_path = classify_put_write_path(is_inline_buffer, put_object_size, fi.erasure.block_size);
             let direct_inline_commit = matches!(write_path, SmallWritePath::Inline);
+            {
+                use std::io::Write;
+                let msg = format!(
+                    "INLINE_DEBUG: bucket={} obj={} size={} shard_fs={} ds={} bs={} inline={} direct={} path={} iblock={} ver={}\n",
+                    bucket, object, put_object_size, shard_file_size_raw, erasure.data_shards, fi.erasure.block_size,
+                    is_inline_buffer, direct_inline_commit, write_path.metric_label(), storage_class_config.inline_block(), opts.versioned
+                );
+                if let Ok(mut f) = std::fs::OpenOptions::new().create(true).append(true).open("/tmp/rustfs_inline_debug.log") {
+                    let _ = f.write_all(msg.as_bytes());
+                }
+                let _ = std::io::stderr().write_all(msg.as_bytes());
+            }
             rustfs_io_metrics::record_put_object_path(write_path.metric_label());
             let writer_setup_stage_start = collect_stage_timing.then(Instant::now);
             let (mut writers, errors) = if direct_inline_commit {
@@ -5655,7 +5669,18 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         objects: Vec<ObjectToDelete>,
         opts: ObjectOptions,
     ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
+        let (deleted, errors, _) = self.delete_objects_with_accounting(bucket, objects, opts).await;
+        (deleted, errors)
+    }
+
+    async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
         let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut accounting = vec![None; objects.len()];
         let delete_config_snapshot = opts
             .delete_replication_config_snapshot
             .clone()
@@ -5745,7 +5770,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                                 *item = Some(Error::other(message.clone()));
                             }
                         }
-                        return (del_objects, del_errs);
+                        return (del_objects, del_errs, accounting);
                     }
                 },
             }
@@ -5792,6 +5817,22 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             let source_missing = gerr
                 .as_ref()
                 .is_some_and(|err| is_err_object_not_found(err) || is_err_version_not_found(err));
+            // Resolve accounting from the generation selected under this
+            // object's write lock. A request-layer pre-stat is only an
+            // optimization and cannot identify a concurrent overwrite.
+            let (accounting_size, accounting_version_id, removed_current_object) = if source_missing
+                || dobj.synthetic_version_id
+                || set_disk_delete_creates_delete_marker(&check_opts)
+                || goi.delete_marker
+            {
+                (None, None, false)
+            } else {
+                (
+                    quota_object_size(&goi).ok(),
+                    goi.version_id.filter(|version_id| !version_id.is_nil()),
+                    (dobj.version_id.is_none() || is_explicit_null_version(dobj.version_id)) && !dobj.synthetic_version_id,
+                )
+            };
             // Normalize both sides before comparing. `goi.version_id` is the
             // client-facing identity, where `from_file_info` synthesizes
             // `Some(Uuid::nil())` for a null version on a versioned or
@@ -5920,7 +5961,12 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                     },
                     replication_state: vr.replication_state_internal.clone(),
                     ..Default::default()
-                }
+                };
+                accounting[i] = Some(DeleteAccounting {
+                    size: accounting_size,
+                    version_id: accounting_version_id,
+                    removed_current_object,
+                });
             }
 
             // Only add to vers_map if we hold the lock
@@ -5966,7 +6012,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                     });
                 }
             }
-            return (del_objects, del_errs);
+            return (del_objects, del_errs, accounting);
         }
 
         let mut persisted_journal_entries = Vec::with_capacity(journal_entries.len());
@@ -6204,7 +6250,16 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             }
         }
 
-        (del_objects, del_errs)
+        // An accounting identity is actionable only when the delete result is
+        // successful. Never let a failed commit (including a partial quorum
+        // failure) reach the request-layer fast delta path.
+        for (index, err) in del_errs.iter().enumerate() {
+            if err.is_some() {
+                accounting[index] = None;
+            }
+        }
+
+        (del_objects, del_errs, accounting)
     }
 
     #[tracing::instrument(skip(self))]
@@ -6533,6 +6588,12 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
 
         let mut obj_info = ObjectInfo::from_file_info(&dfi, bucket, object, opts.versioned || opts.version_suspended);
         obj_info.size = goi.size;
+        // Keep the committed source metadata on the internal delete result so
+        // the request layer can derive canonical accounting for this exact
+        // generation. Delete responses do not expose these fields.
+        obj_info.actual_size = goi.actual_size;
+        obj_info.user_defined = Arc::clone(&goi.user_defined);
+        obj_info.parts = Arc::clone(&goi.parts);
         obj_info.user_tags = Arc::clone(&goi.user_tags);
         self.invalidate_get_object_metadata_cache(bucket, object).await;
         Ok(obj_info)
@@ -7824,6 +7885,113 @@ mod replication_quota_safety_tests {
         assert_eq!(stored.get_actual_size().expect("stored logical size should parse"), 1);
     }
 
+    #[tokio::test]
+    async fn delete_returns_canonical_compressed_accounting_size() {
+        let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "compressed-delete-accounting";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let mut user_defined = HashMap::new();
+        insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        insert_str(&mut user_defined, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+        let mut reader = PutObjReader::new(
+            HashReader::from_stream(Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid"),
+        );
+        set_disks
+            .put_object(
+                bucket,
+                "object",
+                &mut reader,
+                &ObjectOptions {
+                    user_defined,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("compressed object should be written");
+
+        let (deleted, errors, accounting) = set_disks
+            .delete_objects_with_accounting(
+                bucket,
+                vec![ObjectToDelete {
+                    object_name: "object".to_string(),
+                    ..Default::default()
+                }],
+                ObjectOptions {
+                    object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(
+                        ObjectLockConfigState::ConfirmedAbsent,
+                    ))),
+                    ..Default::default()
+                },
+            )
+            .await;
+
+        assert!(errors[0].is_none(), "compressed delete should succeed: {:?}", errors[0]);
+        assert!(deleted[0].found, "the committed object must be reported as found");
+        assert_eq!(accounting[0].as_ref().and_then(|value| value.size), Some(1000));
+        assert!(accounting[0].as_ref().is_some_and(|value| value.version_id.is_none()));
+        assert!(accounting[0].as_ref().is_some_and(|value| value.removed_current_object));
+    }
+
+    #[tokio::test]
+    async fn suspended_delete_marker_does_not_return_body_accounting() {
+        let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "suspended-delete-accounting";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let mut user_defined = HashMap::new();
+        insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        insert_str(&mut user_defined, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+        let mut reader = PutObjReader::new(
+            HashReader::from_stream(Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid"),
+        );
+        let suspended_opts = ObjectOptions {
+            version_suspended: true,
+            delete_replication_config_snapshot: Some(Arc::new(DeleteReplicationConfigSnapshot::from_configs_for_test(
+                s3s::dto::VersioningConfiguration {
+                    status: Some(s3s::dto::BucketVersioningStatus::from_static(s3s::dto::BucketVersioningStatus::SUSPENDED)),
+                    ..Default::default()
+                },
+                None,
+            ))),
+            user_defined,
+            object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(ObjectLockConfigState::ConfirmedAbsent))),
+            ..Default::default()
+        };
+        set_disks
+            .put_object(bucket, "object", &mut reader, &suspended_opts)
+            .await
+            .expect("compressed object should be written");
+
+        let (deleted, errors, accounting) = set_disks
+            .delete_objects_with_accounting(
+                bucket,
+                vec![ObjectToDelete {
+                    object_name: "object".to_string(),
+                    ..Default::default()
+                }],
+                suspended_opts,
+            )
+            .await;
+        assert!(errors[0].is_none(), "suspended delete should create a marker: {:?}", errors[0]);
+        assert!(deleted[0].delete_marker);
+        assert!(accounting[0].is_none(), "a delete marker must not carry body accounting");
+    }
+
     #[tokio::test]
     async fn direct_put_cannot_persist_a_tiny_logical_size() {
         let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;

crates/ecstore/src/storage_api_contracts/mod.rs

@@ -62,8 +62,8 @@ pub(crate) mod object {
     use super::{Debug, Error, FileInfo, GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
     use crate::storage_api_contracts::range::HTTPRangeSpec;
     pub(crate) use rustfs_storage_api::{
-        DeletedObject, HTTPPreconditions, ObjectIO, ObjectLockDeleteOptions, ObjectLockRetentionOptions, ObjectOperations,
-        ObjectPreconditionError, ObjectPreconditionPart, ObjectPreconditionState, ObjectToDelete,
+        DeleteAccounting, DeletedObject, HTTPPreconditions, ObjectIO, ObjectLockDeleteOptions, ObjectLockRetentionOptions,
+        ObjectOperations, ObjectPreconditionError, ObjectPreconditionPart, ObjectPreconditionState, ObjectToDelete,
     };
 
     pub(crate) trait EcstoreObjectIO:

crates/ecstore/src/store/heal.rs

@@ -297,10 +297,16 @@ impl ECStore {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::bucket::metadata_sys;
     use crate::core::pools::{PoolDecommissionInfo, PoolStatus};
-    use crate::disk::{DiskOption, format::FormatV3, new_disk};
-    use crate::layout::endpoints::{Endpoints, PoolEndpoints};
+    use crate::disk::{DeleteOptions, DiskOption, format::FormatV3, new_disk};
+    use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
+    use crate::runtime::instance::InstanceContext;
+    use crate::storage_api_contracts::bucket::{BucketOperations, MakeBucketOptions};
+    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations};
     use crate::store::init_format::{load_format_erasure, save_format_file};
+    use crate::store::init_local_disks_with_instance_ctx;
+    use tokio_util::sync::CancellationToken;
 
     async fn minimal_heal_pool(pool_idx: usize) -> Arc<Sets> {
         let format = FormatV3::new(1, 1);
@@ -347,6 +353,51 @@ mod tests {
         }
     }
 
+    async fn multi_pool_heal_store() -> (tempfile::TempDir, Arc<ECStore>, CancellationToken) {
+        let temp_dir = tempfile::tempdir().expect("multi-pool heal test directory should be created");
+        let mut pool_endpoints = Vec::new();
+        for pool_index in 0..2 {
+            let mut endpoints = Vec::new();
+            for disk_index in 0..4 {
+                let disk_path = temp_dir.path().join(format!("pool{pool_index}-disk{disk_index}"));
+                tokio::fs::create_dir_all(&disk_path)
+                    .await
+                    .expect("multi-pool heal test disk should be created");
+                let mut endpoint = Endpoint::try_from(disk_path.to_str().expect("disk path should be utf8"))
+                    .expect("test endpoint should parse");
+                endpoint.set_pool_index(pool_index);
+                endpoint.set_set_index(0);
+                endpoint.set_disk_index(disk_index);
+                endpoints.push(endpoint);
+            }
+            pool_endpoints.push(PoolEndpoints {
+                legacy: false,
+                set_count: 1,
+                drives_per_set: 4,
+                endpoints: Endpoints::from(endpoints),
+                cmd_line: format!("heal-owner-pool-{pool_index}"),
+                platform: "test".to_string(),
+            });
+        }
+
+        let endpoint_pools = EndpointServerPools::from(pool_endpoints);
+        let instance_ctx = Arc::new(InstanceContext::new());
+        init_local_disks_with_instance_ctx(&instance_ctx, endpoint_pools.clone())
+            .await
+            .expect("multi-pool local disks should initialize");
+        let shutdown = CancellationToken::new();
+        let store = ECStore::new_with_instance_ctx(
+            "127.0.0.1:0".parse().expect("test address should parse"),
+            endpoint_pools,
+            shutdown.clone(),
+            instance_ctx,
+        )
+        .await
+        .expect("multi-pool test store should initialize");
+        metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        (temp_dir, store, shutdown)
+    }
+
     #[tokio::test]
     async fn heal_object_pool_scope_selects_only_requested_pool() {
         let store = minimal_heal_store().await;
@@ -506,6 +557,229 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn unscoped_heal_object_suspended_owner_semantics() {
+        let (_temp_dir, store, shutdown) = multi_pool_heal_store().await;
+        let bucket = format!("heal-owner-{}", Uuid::new_v4().simple());
+        let active_object = "active-owner";
+        let suspended_only_object = "suspended-only";
+        let duplicate_object = "duplicate-owner";
+        let marker_object = "marker-owner";
+        let quorum_object = "quorum-owner";
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created in all pools");
+
+        let mut active_reader = PutObjReader::from_vec(b"active owner".to_vec());
+        store.pools[0]
+            .put_object(&bucket, active_object, &mut active_reader, &ObjectOptions::default())
+            .await
+            .expect("active owner object should be written");
+        let active_disks = store.pools[0].disk_set[0].disks.read().await.clone();
+        let missing_active_disk = active_disks[0].clone().expect("active disk should be online");
+        missing_active_disk
+            .delete(
+                &bucket,
+                active_object,
+                DeleteOptions {
+                    recursive: true,
+                    immediate: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("active owner shard should be removed for repair");
+        assert!(
+            missing_active_disk.read_xl(&bucket, active_object, false).await.is_err(),
+            "the active owner fixture must start with one missing metadata copy"
+        );
+
+        let mut suspended_reader = PutObjReader::from_vec(b"suspended owner".to_vec());
+        store.pools[1]
+            .put_object(&bucket, suspended_only_object, &mut suspended_reader, &ObjectOptions::default())
+            .await
+            .expect("suspended owner object should be written");
+        for (pool_index, mod_time) in [1_i64, 2_i64].into_iter().enumerate() {
+            let mut duplicate_reader = PutObjReader::from_vec(format!("duplicate-pool-{pool_index}").into_bytes());
+            store.pools[pool_index]
+                .put_object(
+                    &bucket,
+                    duplicate_object,
+                    &mut duplicate_reader,
+                    &ObjectOptions {
+                        mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(mod_time)),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("duplicate owner object should be written");
+        }
+        let duplicate_missing_disk = store.pools[0].disk_set[0].disks.read().await[0]
+            .clone()
+            .expect("duplicate active owner disk should be online");
+        duplicate_missing_disk
+            .delete(
+                &bucket,
+                duplicate_object,
+                DeleteOptions {
+                    recursive: true,
+                    immediate: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("duplicate active owner shard should be removed for repair");
+        let history_version = Uuid::new_v4();
+        let mut history_reader = PutObjReader::from_vec(b"marker history".to_vec());
+        store.pools[0]
+            .put_object(
+                &bucket,
+                marker_object,
+                &mut history_reader,
+                &ObjectOptions {
+                    versioned: true,
+                    version_id: Some(history_version.to_string()),
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(1)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("versioned marker history should be written");
+        store.pools[0]
+            .delete_object(
+                &bucket,
+                marker_object,
+                ObjectOptions {
+                    versioned: true,
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(2)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("delete marker should be written");
+        let mut quorum_reader = PutObjReader::from_vec(b"quorum boundary".to_vec());
+        store.pools[0]
+            .put_object(&bucket, quorum_object, &mut quorum_reader, &ObjectOptions::default())
+            .await
+            .expect("quorum boundary object should be written");
+        {
+            let mut pool_meta = store.pool_meta.write().await;
+            let mut next = PoolMeta::new(&store.pools, &pool_meta);
+            next.pools[1].decommission = Some(PoolDecommissionInfo {
+                start_time: Some(OffsetDateTime::UNIX_EPOCH),
+                ..Default::default()
+            });
+            *pool_meta = next;
+        }
+
+        let (_, duplicate_owner) = store
+            .get_latest_object_info_with_idx(&bucket, duplicate_object, &ObjectOptions::default())
+            .await
+            .expect("duplicate owner should resolve");
+        assert_eq!(duplicate_owner, 1, "latest duplicate must win when all pools are eligible");
+        let (_, active_duplicate_owner) = store
+            .get_latest_object_info_with_idx(
+                &bucket,
+                duplicate_object,
+                &ObjectOptions {
+                    skip_decommissioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("active duplicate owner should resolve");
+        assert_eq!(
+            active_duplicate_owner, 0,
+            "suspended duplicate must be excluded from active owner selection"
+        );
+        let (duplicate_result, duplicate_err) = store
+            .handle_heal_object(&bucket, duplicate_object, "", &HealOpts::default())
+            .await
+            .expect("duplicate owner heal should complete through the production path");
+        assert_eq!(duplicate_result.object, duplicate_object);
+        assert!(duplicate_err.is_none(), "active duplicate should be repaired: {duplicate_err:?}");
+        assert!(
+            duplicate_missing_disk.read_xl(&bucket, duplicate_object, false).await.is_ok(),
+            "production heal must repair the active duplicate owner rather than the suspended owner"
+        );
+        let (marker_info, marker_owner) = store
+            .get_latest_object_info_with_idx(
+                &bucket,
+                marker_object,
+                &ObjectOptions {
+                    skip_decommissioned: true,
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("latest delete marker should resolve");
+        assert_eq!(marker_owner, 0);
+        assert!(marker_info.delete_marker, "latest version must preserve delete-marker semantics");
+
+        let (active_result, active_err) = store
+            .handle_heal_object(&bucket, active_object, "", &HealOpts::default())
+            .await
+            .expect("unscoped active-owner heal should complete");
+        assert_eq!(active_result.object, active_object);
+        assert!(active_err.is_none(), "active owner must be selected even with a suspended pool");
+        assert!(
+            missing_active_disk.read_xl(&bucket, active_object, false).await.is_ok(),
+            "active owner heal must write the missing disk metadata: result={active_result:?}, err={active_err:?}"
+        );
+        assert!(
+            store.pools[1]
+                .get_object_info(&bucket, active_object, &ObjectOptions::default())
+                .await
+                .is_err(),
+            "the suspended pool must not be written for an active-owner object"
+        );
+
+        let (suspended_result, suspended_err) = store
+            .handle_heal_object(&bucket, suspended_only_object, "", &HealOpts::default())
+            .await
+            .expect("unscoped suspended-only heal should return a terminal result");
+        assert!(suspended_result.object.is_empty());
+        assert!(matches!(suspended_err, Some(Error::FileNotFound)));
+        assert!(
+            store.pools[1]
+                .get_object_info(&bucket, suspended_only_object, &ObjectOptions::default())
+                .await
+                .is_ok(),
+            "suspended-only data must remain untouched when unscoped heal reports absent"
+        );
+
+        let (_, explicit_err) = store
+            .handle_heal_object(
+                &bucket,
+                suspended_only_object,
+                "",
+                &HealOpts {
+                    pool: Some(1),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("explicit suspended-owner heal should return a mapped error");
+        assert!(matches!(explicit_err, Some(Error::SlowDown)));
+
+        let original_quorum_disks = store.pools[0].disk_set[0].disks.read().await.clone();
+        let surviving_quorum_disk = original_quorum_disks[3].clone();
+        *store.pools[0].disk_set[0].disks.write().await = vec![None, None, None, surviving_quorum_disk];
+        let (_, quorum_err) = store
+            .handle_heal_object(&bucket, quorum_object, "", &HealOpts::default())
+            .await
+            .expect("quorum boundary heal should return a mapped result");
+        *store.pools[0].disk_set[0].disks.write().await = original_quorum_disks;
+        assert!(
+            matches!(quorum_err, Some(Error::ErasureReadQuorum)),
+            "quorum-boundary heal must preserve quorum error, got {quorum_err:?}"
+        );
+        shutdown.cancel();
+    }
+
     #[tokio::test]
     async fn handle_heal_format_continues_after_a_pool_error() {
         let canonical_format = FormatV3::new(1, 3);

crates/ecstore/src/store/object.rs

@@ -41,7 +41,7 @@ use crate::set_disk::{
 };
 use crate::storage_api_contracts::{
     namespace::NamespaceLocking as _,
-    object::{ObjectIO as _, ObjectOperations as _},
+    object::{DeleteAccounting, ObjectIO as _, ObjectOperations as _},
 };
 use parking_lot::Mutex as ParkingMutex;
 use rustfs_io_metrics::{
@@ -1216,6 +1216,14 @@ fn return_batch_delete_lock_error(objects: &[ObjectToDelete], err: Error) -> (Ve
     (del_objects, del_errs)
 }
 
+fn return_batch_delete_lock_error_with_accounting(
+    objects: &[ObjectToDelete],
+    err: Error,
+) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+    let (deleted, errors) = return_batch_delete_lock_error(objects, err);
+    (deleted, errors, vec![None; objects.len()])
+}
+
 fn sorted_unique_delete_object_names(objects: &[ObjectToDelete]) -> Vec<&str> {
     let mut object_names: Vec<&str> = objects.iter().map(|object| object.object_name.as_str()).collect();
     object_names.sort_unstable();
@@ -2312,6 +2320,22 @@ impl ECStore {
         result
     }
 
+    pub async fn delete_objects_with_tier_delete_journal_and_accounting(
+        self: &Arc<Self>,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+        let result = self
+            .handle_delete_objects_with_journal_and_accounting(bucket, objects, opts, Some(Arc::clone(self)))
+            .await;
+        let success_count = result.1.iter().filter(|err| err.is_none()).count();
+        if success_count > 0 {
+            list_objects::observe_list_objects_mutations(self, bucket, success_count).await;
+        }
+        result
+    }
+
     #[instrument(skip(self))]
     pub(super) async fn handle_delete_object(&self, bucket: &str, object: &str, opts: ObjectOptions) -> Result<ObjectInfo> {
         self.handle_delete_object_with_journal(bucket, object, opts, None).await
@@ -2689,6 +2713,19 @@ impl ECStore {
         opts: ObjectOptions,
         tier_journal_api: Option<Arc<ECStore>>,
     ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
+        let (deleted, errors, _) = self
+            .handle_delete_objects_with_journal_and_accounting(bucket, objects, opts, tier_journal_api)
+            .await;
+        (deleted, errors)
+    }
+
+    pub(super) async fn handle_delete_objects_with_journal_and_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+        tier_journal_api: Option<Arc<ECStore>>,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
         // encode object name
         let objects: Vec<ObjectToDelete> = objects
             .iter()
@@ -2701,6 +2738,7 @@ impl ECStore {
 
         // Default return value
         let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut accounting = vec![None; objects.len()];
 
         let mut del_errs = Vec::with_capacity(objects.len());
         for _ in 0..objects.len() {
@@ -2714,7 +2752,7 @@ impl ECStore {
         } else {
             match self.acquire_bucket_lifecycle_read_lock(bucket).await {
                 Ok(guard) => Some(guard),
-                Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
             }
         };
         if let Some(guard) = _bucket_lifecycle_guard.as_ref() {
@@ -2726,21 +2764,21 @@ impl ECStore {
                 Err(err) => {
                     let message = err.to_string();
                     let errors = (0..objects.len()).map(|_| Some(Error::other(message.clone()))).collect();
-                    return (del_objects, errors);
+                    return (del_objects, errors, accounting);
                 }
             }
         }
         if !is_meta_bucketname(bucket)
             && let Err(err) = get_cached_bucket_incarnation_id_in(&self.ctx, bucket).await
         {
-            return return_batch_delete_lock_error(objects.as_slice(), err);
+            return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err);
         }
         let _object_lock_metadata_guard = if is_meta_bucketname(bucket) {
             None
         } else {
             Some(match acquire_bucket_metadata_transaction_read_lock_in(&self.ctx, bucket).await {
                 Ok(guard) => guard,
-                Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
             })
         };
         if let Some(guard) = _object_lock_metadata_guard.as_ref() {
@@ -2750,7 +2788,7 @@ impl ECStore {
             let (state, incarnation_id, config_revision) =
                 match get_object_lock_config_and_incarnation_from_disk_in(&self.ctx, bucket).await {
                     Ok(snapshot) => snapshot,
-                    Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                    Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
                 };
             opts.object_lock_config_snapshot = Some(Arc::new(ObjectLockConfigSnapshot::for_store_bucket(
                 self.id,
@@ -2766,7 +2804,10 @@ impl ECStore {
         if let (Some(expected), Some(current)) = (opts.expected_bucket_incarnation_id, current_bucket_incarnation_id)
             && expected != current
         {
-            return return_batch_delete_lock_error(objects.as_slice(), StorageError::BucketNotFound(bucket.to_string()));
+            return return_batch_delete_lock_error_with_accounting(
+                objects.as_slice(),
+                StorageError::BucketNotFound(bucket.to_string()),
+            );
         }
         #[cfg(test)]
         if current_bucket_incarnation_id.is_some() {
@@ -2774,7 +2815,7 @@ impl ECStore {
         }
         let _object_lock_guards = match self.acquire_delete_objects_write_locks(bucket, &objects, &mut opts).await {
             Ok(guards) => guards,
-            Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+            Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
         };
 
         let mut futures = Vec::with_capacity(self.pools.len());
@@ -2783,22 +2824,24 @@ impl ECStore {
             if self.is_pool_rebalancing(pool.pool_idx).await {
                 continue;
             }
-            futures.push(pool.delete_objects(bucket, objects.clone(), opts.clone()));
+            futures.push(pool.delete_objects_with_accounting(bucket, objects.clone(), opts.clone()));
         }
 
         let results = join_all(futures).await;
 
         for idx in 0..del_objects.len() {
-            for (dels, errs) in results.iter() {
+            for (dels, errs, pool_accounting) in results.iter() {
                 if errs[idx].is_none() && dels[idx].found {
                     del_errs[idx] = None;
                     del_objects[idx] = dels[idx].clone();
+                    accounting[idx] = pool_accounting[idx].clone();
                     break;
                 }
 
                 if del_errs[idx].is_none() {
                     del_errs[idx] = errs[idx].clone();
                     del_objects[idx] = dels[idx].clone();
+                    accounting[idx] = pool_accounting[idx].clone();
                 }
             }
         }
@@ -2807,7 +2850,7 @@ impl ECStore {
             v.object_name = decode_dir_object(&v.object_name);
         });
 
-        (del_objects, del_errs)
+        (del_objects, del_errs, accounting)
 
         // let mut futures = Vec::with_capacity(objects.len());
 

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/heal/Cargo.toml

@@ -91,6 +91,7 @@ metrics = { workspace = true }
 base64 = { workspace = true }
 bytes = { workspace = true }
 crc-fast = { workspace = true }
+sha2 = { workspace = true }
 
 [dev-dependencies]
 serde_json = { workspace = true, features = ["raw_value"] }

crates/heal/src/heal/erasure_healer.rs

@@ -373,6 +373,11 @@ impl ErasureSetHealer {
         set_disk_id: &str,
         buckets: &[String],
     ) -> Result<(ResumeManager, CheckpointManager)> {
+        if self.replacement_task_id.is_none() && CheckpointManager::is_blocked(&self.disk, task_id).await {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Resume task {task_id} has a blocked checkpoint"),
+            });
+        }
         // check if resume state exists
         let has_resume_state = if self.replacement_task_id.is_some() {
             ResumeManager::has_replacement_intent(&self.disk, task_id).await

crates/heal/src/heal/resume.rs

@@ -51,6 +51,7 @@ const RESUME_STATE_FILE: &str = "ahm_resume_state.json";
 const REPLACEMENT_INTENT_FILE: &str = "ahm_replacement_intent.json";
 const RESUME_PROGRESS_FILE: &str = "ahm_progress.json";
 pub(super) const RESUME_CHECKPOINT_FILE: &str = "ahm_checkpoint.json";
+pub(super) const RESUME_CHECKPOINT_BLOCKED_FILE: &str = "ahm_checkpoint.blocked";
 const REPLACEMENT_COMPLETION_PROOF_FILE: &str = "ahm_replacement_completion_proof.json";
 const REPLACEMENT_RECOVERY_DIR: &str = "ahm-replacement";
 const REPLACEMENT_INTENT_SEAL_FILE: &str = "ahm_replacement_intent_seal";

crates/heal/src/heal/resume/checkpoint.rs

@@ -13,26 +13,31 @@
 // limitations under the License.
 
 use crate::{Error, Result};
+use base64::Engine as _;
 use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
 use std::collections::HashSet;
 use std::path::Path;
 use std::sync::{Arc, Mutex};
 use std::time::{SystemTime, UNIX_EPOCH};
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex as AsyncMutex, RwLock};
 use tracing::{debug, warn};
 
-use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
+use super::super::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskAPI, EcstoreDiskBytes};
+use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt, RUSTFS_META_BUCKET};
 use super::{
-    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_FILE, delete_resume_file, path_to_str,
-    validate_resume_task_id,
+    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_BLOCKED_FILE, RESUME_CHECKPOINT_FILE,
+    delete_resume_file, path_to_str, validate_resume_task_id,
 };
 
 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
+const RESUME_CHECKPOINT_DIGEST_FILE: &str = "ahm_checkpoint.sha256";
+const CHECKPOINT_PER_VERSION_SCHEMA: u32 = 5;
 
 /// Current on-disk schema version for `ResumeCheckpoint`. Same rationale as
 /// `CURRENT_RESUME_SCHEMA`: pre-per-version dedup identities are not comparable
 /// to the new `compose_key` identities, so a stale checkpoint is discarded.
-pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
+pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 6;
 
 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -57,6 +62,11 @@ pub struct ResumeCheckpoint {
     pub failed_objects: HashSet<String>,
     /// skipped objects
     pub skipped_objects: HashSet<String>,
+    /// Integrity digest over the checkpoint with this field set to `None`.
+    /// Keeping it in the checkpoint makes the payload and its authentication
+    /// record one CAS generation instead of two independently-written files.
+    #[serde(default)]
+    pub integrity_digest: Option<String>,
 }
 
 impl ResumeCheckpoint {
@@ -70,6 +80,7 @@ impl ResumeCheckpoint {
             processed_objects: HashSet::new(),
             failed_objects: HashSet::new(),
             skipped_objects: HashSet::new(),
+            integrity_digest: None,
         }
     }
 
@@ -116,17 +127,111 @@ pub struct CheckpointManager {
     disk: DiskStore,
     checkpoint: Arc<RwLock<ResumeCheckpoint>>,
     throttle: Mutex<PersistThrottle>,
+    save_lock: AsyncMutex<()>,
+    last_saved: Mutex<Option<EcstoreDiskBytes>>,
 }
 
 impl CheckpointManager {
+    fn blocked_path(task_id: &str) -> std::path::PathBuf {
+        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}"))
+    }
+
+    /// Return whether a checkpoint was permanently isolated after a malformed
+    /// or unsupported snapshot was observed.
+    pub(crate) async fn is_blocked(disk: &DiskStore, task_id: &str) -> bool {
+        if validate_resume_task_id(task_id).is_err() {
+            return false;
+        }
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return false;
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(_) => true,
+            Err(crate::heal::DiskError::FileNotFound) => false,
+            Err(_) => true,
+        }
+    }
+
+    /// Validate the checkpoint while enumerating resumable state. This reads
+    /// the checkpoint once and also isolates malformed or unsupported data.
+    pub(crate) async fn is_resumable(disk: &DiskStore, task_id: &str) -> Result<bool> {
+        validate_resume_task_id(task_id)?;
+        if Self::is_blocked(disk, task_id).await {
+            return Err(Error::InvalidCheckpoint(format!("Resume task {task_id} has a blocked checkpoint")));
+        }
+        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
+        let Ok(path) = path_to_str(&file_path) else {
+            return Err(Error::InvalidCheckpoint("Resume checkpoint path is not valid UTF-8".to_string()));
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(bytes) if bytes.is_empty() => Ok(true),
+            Ok(bytes) => Self::load_from_data(disk.clone(), task_id, bytes.to_vec())
+                .await
+                .map(|_| true),
+            Err(crate::heal::DiskError::FileNotFound) => Ok(true),
+            Err(error) => Err(error.into()),
+        }
+    }
+
+    async fn block_invalid_snapshot(disk: &DiskStore, task_id: &str) {
+        // This marker is intentionally version-agnostic: an unsupported reader
+        // must stop selector retries until an operator cleans up the snapshot.
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return;
+        };
+        let result = EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path,
+            None,
+            Some(EcstoreDiskBytes::from_static(b"blocked")),
+        )
+        .await;
+        match result {
+            Ok(EcstoreConditionalFileUpdate::Updated | EcstoreConditionalFileUpdate::Mismatch) => {}
+            Ok(EcstoreConditionalFileUpdate::Missing) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = "marker target disappeared",
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+            Err(error) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = %error,
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+        }
+    }
+
     /// create new checkpoint manager
     pub async fn new(disk: DiskStore, task_id: String) -> Result<Self> {
         validate_resume_task_id(&task_id)?;
+        let checkpoint_volume = format!("{RUSTFS_META_BUCKET}/{BUCKET_META_PREFIX}");
+        if let Err(error) = EcstoreDiskAPI::make_volume(disk.as_ref(), &checkpoint_volume).await
+            && error != crate::heal::DiskError::VolumeExists
+        {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Failed to create checkpoint volume: {error}"),
+            });
+        }
         let checkpoint = ResumeCheckpoint::new(task_id);
         let manager = Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(None),
         };
 
         // save initial checkpoint
@@ -140,6 +245,7 @@ impl CheckpointManager {
                 error = %e,
                 "Heal checkpoint persistence failed"
             );
+            return Err(e);
         }
         Ok(manager)
     }
@@ -148,11 +254,22 @@ impl CheckpointManager {
     pub async fn load_from_disk(disk: DiskStore, task_id: &str) -> Result<Self> {
         validate_resume_task_id(task_id)?;
         let checkpoint_data = Self::read_checkpoint_file(&disk, task_id).await?;
-        let mut checkpoint: ResumeCheckpoint =
-            serde_json::from_slice(&checkpoint_data).map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to deserialize checkpoint: {e}"),
-            })?;
+        Self::load_from_data(disk, task_id, checkpoint_data).await
+    }
+
+    async fn load_from_data(disk: DiskStore, task_id: &str, checkpoint_data: Vec<u8>) -> Result<Self> {
+        validate_resume_task_id(task_id)?;
+        let mut checkpoint: ResumeCheckpoint = match serde_json::from_slice(&checkpoint_data) {
+            Ok(checkpoint) => checkpoint,
+            Err(error) => {
+                Self::block_invalid_snapshot(&disk, task_id).await;
+                return Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to deserialize checkpoint: {error}"),
+                });
+            }
+        };
         if checkpoint.task_id != task_id {
+            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: "Resume checkpoint task id does not match filename".to_string(),
             });
@@ -163,6 +280,7 @@ impl CheckpointManager {
         // identities. Discard the stale sets and position, then stamp the
         // current schema so the scan restarts cleanly.
         if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+            Self::block_invalid_snapshot(&disk, task_id).await;
             return Err(Error::TaskExecutionFailed {
                 message: format!(
                     "Checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
@@ -170,7 +288,43 @@ impl CheckpointManager {
                 ),
             });
         }
-        if checkpoint.schema_version < CURRENT_CHECKPOINT_SCHEMA {
+
+        if let Some(expected) = checkpoint.integrity_digest.as_deref() {
+            let actual = Self::checkpoint_digest(&Self::serialize_without_digest(&checkpoint)?);
+            if expected != actual {
+                Self::block_invalid_snapshot(&disk, task_id).await;
+                return Err(Error::InvalidCheckpoint(format!(
+                    "Resume checkpoint digest does not match task {task_id}"
+                )));
+            }
+        } else if checkpoint.schema_version >= CURRENT_CHECKPOINT_SCHEMA {
+            Self::block_invalid_snapshot(&disk, task_id).await;
+            return Err(Error::InvalidCheckpoint(format!(
+                "Resume checkpoint digest is missing for task {task_id}"
+            )));
+        } else {
+            let digest_path = Self::digest_path(task_id);
+            let digest_path = path_to_str(&digest_path)?;
+            match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, digest_path).await {
+                Ok(expected) => {
+                    let actual = Self::checkpoint_digest(&checkpoint_data);
+                    if expected.as_ref() != actual.as_bytes() {
+                        Self::block_invalid_snapshot(&disk, task_id).await;
+                        return Err(Error::InvalidCheckpoint(format!(
+                            "Resume checkpoint digest does not match task {task_id}"
+                        )));
+                    }
+                }
+                Err(crate::heal::DiskError::FileNotFound) => {}
+                Err(error) => {
+                    return Err(Error::TaskExecutionFailed {
+                        message: format!("Failed to read checkpoint digest: {error}"),
+                    });
+                }
+            }
+        }
+
+        if checkpoint.schema_version < CHECKPOINT_PER_VERSION_SCHEMA {
             warn!(
                 target: "rustfs::heal::resume",
                 event = EVENT_HEAL_CHECKPOINT_STATE,
@@ -187,13 +341,15 @@ impl CheckpointManager {
             checkpoint.skipped_objects.clear();
             checkpoint.current_bucket_index = 0;
             checkpoint.current_object_index = 0;
-            checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
         }
+        checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
 
         Ok(Self {
             disk,
             checkpoint: Arc::new(RwLock::new(checkpoint)),
             throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(Some(EcstoreDiskBytes::from(checkpoint_data))),
         })
     }
 
@@ -204,7 +360,7 @@ impl CheckpointManager {
         }
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         match path_to_str(&file_path) {
-            Ok(path_str) => match disk.read_all(RUSTFS_META_BUCKET, path_str).await {
+            Ok(path_str) => match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
                 Ok(data) => !data.is_empty(),
                 Err(_) => false,
             },
@@ -292,6 +448,8 @@ impl CheckpointManager {
 
         let checkpoint_file = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
         delete_resume_file(&self.disk, &checkpoint_file).await?;
+        delete_resume_file(&self.disk, &Self::digest_path(&task_id)).await?;
+        delete_resume_file(&self.disk, &Self::blocked_path(&task_id)).await?;
 
         debug!(
             target: "rustfs::heal::resume",
@@ -307,21 +465,130 @@ impl CheckpointManager {
 
     /// save checkpoint to disk
     async fn save_checkpoint(&self) -> Result<()> {
-        let checkpoint = self.checkpoint.read().await;
+        // Serialize saves and take the snapshot only after acquiring the lock:
+        // a slower writer must not publish a snapshot taken before a newer one.
+        let _save_guard = self.save_lock.lock().await;
+        let checkpoint = self.checkpoint.read().await.clone();
         validate_resume_task_id(&checkpoint.task_id)?;
-        let checkpoint_data = serde_json::to_vec(&*checkpoint).map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to serialize checkpoint: {e}"),
-        })?;
+        let unsigned_checkpoint_data = Self::serialize_without_digest(&checkpoint)?;
+        let digest = Self::checkpoint_digest(&unsigned_checkpoint_data);
+        let mut persisted_checkpoint = checkpoint.clone();
+        persisted_checkpoint.integrity_digest = Some(digest);
+        let checkpoint_data =
+            EcstoreDiskBytes::from(serde_json::to_vec(&persisted_checkpoint).map_err(|e| Error::TaskExecutionFailed {
+                message: format!("Failed to serialize checkpoint: {e}"),
+            })?);
 
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{}_{}", checkpoint.task_id, RESUME_CHECKPOINT_FILE));
 
         let path_str = path_to_str(&file_path)?;
-        self.disk
-            .write_all(RUSTFS_META_BUCKET, path_str, checkpoint_data.into())
+        let last_saved = self
+            .last_saved
+            .lock()
+            .map_err(|_| Error::TaskExecutionFailed {
+                message: "Checkpoint save state lock is poisoned; refusing to save".to_string(),
+            })?
+            .clone();
+        let update = EcstoreDiskAPI::compare_and_update_file(
+            self.disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path_str,
+            last_saved.clone(),
+            Some(checkpoint_data.clone()),
+        )
+        .await
+        .map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to save checkpoint: {e}"),
+        })?;
+
+        let expected = match update {
+            EcstoreConditionalFileUpdate::Updated => None,
+            EcstoreConditionalFileUpdate::Missing => {
+                return Err(Error::TaskExecutionFailed {
+                    message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                });
+            }
+            EcstoreConditionalFileUpdate::Mismatch => {
+                // A healthy manager normally completes the CAS above without
+                // another read or JSON parse. Inspect only after a mismatch so
+                // corruption and future schemas cannot be overwritten blindly.
+                let existing = match HealDiskExt::read_all(self.disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
+                    Ok(existing) => existing,
+                    Err(crate::heal::DiskError::FileNotFound) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                        });
+                    }
+                    Err(error) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!("Failed to inspect checkpoint after CAS mismatch: {error}"),
+                        });
+                    }
+                };
+
+                if existing.is_empty() && last_saved.is_none() {
+                    Some(existing)
+                } else {
+                    let current: ResumeCheckpoint = match serde_json::from_slice(&existing) {
+                        Ok(current) => current,
+                        Err(error) => {
+                            Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                            return Err(Error::TaskExecutionFailed {
+                                message: format!("Existing checkpoint is corrupt: {error}"),
+                            });
+                        }
+                    };
+                    if current.task_id != checkpoint.task_id {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Existing checkpoint task id does not match filename".to_string(),
+                        });
+                    }
+                    if current.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!(
+                                "Existing checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
+                                current.schema_version
+                            ),
+                        });
+                    }
+                    if last_saved.as_ref().is_none_or(|saved| saved.as_ref() != existing.as_ref()) {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint changed since this manager loaded it; refusing to overwrite newer progress"
+                                .to_string(),
+                        });
+                    }
+                    Some(existing)
+                }
+            }
+        };
+
+        if let Some(expected) = expected {
+            match EcstoreDiskAPI::compare_and_update_file(
+                self.disk.as_ref(),
+                RUSTFS_META_BUCKET,
+                path_str,
+                Some(expected),
+                Some(checkpoint_data.clone()),
+            )
             .await
             .map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to save checkpoint: {e}"),
-            })?;
+                message: format!("Failed to save checkpoint after CAS mismatch: {e}"),
+            })? {
+                EcstoreConditionalFileUpdate::Updated => {}
+                EcstoreConditionalFileUpdate::Missing | EcstoreConditionalFileUpdate::Mismatch => {
+                    return Err(Error::TaskExecutionFailed {
+                        message: "Checkpoint changed while saving; refusing to overwrite newer progress".to_string(),
+                    });
+                }
+            }
+        }
+
+        let mut last_saved = self.last_saved.lock().map_err(|_| Error::TaskExecutionFailed {
+            message: "Checkpoint save state lock is poisoned after save".to_string(),
+        })?;
+        *last_saved = Some(checkpoint_data);
 
         debug!(
             target: "rustfs::heal::resume",
@@ -341,11 +608,38 @@ impl CheckpointManager {
         let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
 
         let path_str = path_to_str(&file_path)?;
-        disk.read_all(RUSTFS_META_BUCKET, path_str)
+        HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str)
             .await
             .map(|bytes| bytes.to_vec())
             .map_err(|e| Error::TaskExecutionFailed {
                 message: format!("Failed to read checkpoint file: {e}"),
             })
     }
+
+    fn serialize_without_digest(checkpoint: &ResumeCheckpoint) -> Result<Vec<u8>> {
+        let mut unsigned = checkpoint.clone();
+        unsigned.integrity_digest = None;
+        let mut value = serde_json::to_value(&unsigned).map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to serialize checkpoint: {e}"),
+        })?;
+        for field in ["processed_objects", "failed_objects", "skipped_objects"] {
+            let Some(values) = value.get_mut(field).and_then(serde_json::Value::as_array_mut) else {
+                return Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to canonicalize checkpoint field: {field}"),
+                });
+            };
+            values.sort_by(|left, right| left.as_str().cmp(&right.as_str()));
+        }
+        serde_json::to_vec(&value).map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to serialize checkpoint: {e}"),
+        })
+    }
+
+    fn checkpoint_digest(checkpoint_data: &[u8]) -> String {
+        base64::engine::general_purpose::STANDARD.encode(Sha256::digest(checkpoint_data))
+    }
+
+    fn digest_path(task_id: &str) -> std::path::PathBuf {
+        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_DIGEST_FILE}"))
+    }
 }

crates/heal/src/heal/resume/tests.rs

@@ -1600,6 +1600,29 @@ async fn test_checkpoint_schema_v4_discarded_on_load() {
     temp_dir.close().expect("remove schema test directory");
 }
 
+#[tokio::test]
+async fn unsigned_previous_checkpoint_schema_preserves_progress() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let mut legacy = ResumeCheckpoint::new(task_id.clone());
+    legacy.schema_version = CURRENT_CHECKPOINT_SCHEMA - 1;
+    legacy.update_position(2, 500);
+    legacy.add_processed_object("object".to_string());
+    legacy.integrity_digest = None;
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&legacy).unwrap().into())
+        .await
+        .expect("write previous-schema checkpoint");
+
+    let manager = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    let checkpoint = manager.get_checkpoint().await;
+    assert_eq!(checkpoint.schema_version, CURRENT_CHECKPOINT_SCHEMA);
+    assert_eq!(checkpoint.current_bucket_index, 2);
+    assert_eq!(checkpoint.current_object_index, 500);
+    assert!(checkpoint.processed_objects.contains("object"));
+    temp_dir.close().unwrap();
+}
+
 #[tokio::test]
 async fn current_normal_resume_schema_preserves_progress() {
     let (temp_dir, disk) = schema_test_disk().await;
@@ -1675,6 +1698,369 @@ async fn future_resume_and_checkpoint_schemas_are_rejected() {
     temp_dir.close().expect("remove schema test directory");
 }
 
+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_non_empty_truncated_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let truncated = b"{\"schema_version\":5,\"task_id\":";
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, truncated.as_slice().into())
+        .await
+        .expect("write truncated checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a truncated checkpoint must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint is corrupt"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read truncated checkpoint fixture"),
+        truncated.as_slice()
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove checkpoint save test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_future_schema_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(task_id.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a future schema must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint schema"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read future checkpoint fixture"),
+        future_bytes
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove future schema test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_digest_rejects_same_length_progress_tampering() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    manager
+        .add_processed_object("victim-a".to_string())
+        .await
+        .expect("persist checkpoint progress");
+    manager.update_position(1, 1).await.expect("flush checkpoint progress");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let original = disk
+        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+        .await
+        .expect("read checkpoint fixture");
+    let tampered = original
+        .windows(b"victim-a".len())
+        .position(|window| window == b"victim-a")
+        .map(|index| {
+            let mut bytes = original.to_vec();
+            bytes[index..index + b"victim-a".len()].copy_from_slice(b"victim-b");
+            bytes
+        })
+        .expect("checkpoint should contain the processed object");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, tampered.into())
+        .await
+        .expect("write tampered checkpoint fixture");
+
+    assert!(CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err());
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove digest test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_survives_missing_legacy_sidecar() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    manager.update_position(2, 9).await.unwrap();
+
+    let digest_path = format!("{BUCKET_META_PREFIX}/{task_id}_ahm_checkpoint.sha256");
+    delete_resume_file(&disk, Path::new(&digest_path)).await.unwrap();
+
+    let restored = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    let checkpoint = restored.get_checkpoint().await;
+    assert_eq!(checkpoint.current_bucket_index, 2);
+    assert_eq!(checkpoint.current_object_index, 9);
+    assert!(checkpoint.integrity_digest.is_some());
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_survives_multi_object_reload() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    for index in 0..32 {
+        manager.add_processed_object(format!("processed-{index}")).await.unwrap();
+        manager.add_failed_object(format!("failed-{index}")).await.unwrap();
+        manager.add_skipped_object(format!("skipped-{index}")).await.unwrap();
+    }
+    manager.update_position(2, 9).await.unwrap();
+
+    CheckpointManager::load_from_disk(disk, &task_id)
+        .await
+        .expect("a healthy multi-object checkpoint must survive reload");
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_rejects_a_removed_embedded_digest() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    manager.update_position(2, 9).await.unwrap();
+
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let bytes = disk
+        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+        .await
+        .expect("read checkpoint fixture");
+    let mut value: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
+    value["current_object_index"] = serde_json::json!(10);
+    value.as_object_mut().unwrap().remove("integrity_digest");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&value).unwrap().into())
+        .await
+        .expect("write tampered checkpoint fixture");
+
+    assert!(
+        CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err(),
+        "a current checkpoint without its embedded digest must fail closed"
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn new_checkpoint_manager_rebuilds_an_empty_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty checkpoint fixture");
+
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("a new manager must rebuild an empty checkpoint");
+    manager
+        .update_position(3, 11)
+        .await
+        .expect("rebuilt checkpoint must remain writable");
+    assert!(CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty checkpoint test directory");
+}
+
+#[tokio::test]
+async fn deleted_checkpoint_is_not_recreated_by_an_old_manager() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    manager.cleanup().await.expect("delete checkpoint fixture");
+
+    let error = manager
+        .update_position(1, 2)
+        .await
+        .expect_err("an old manager must not resurrect a deleted checkpoint");
+    assert!(error.to_string().contains("removed after this manager saved it"));
+    assert!(!CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove deleted checkpoint test directory");
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn checkpoint_cleanup_leaves_no_task_specific_lock_artifact() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let lock_path = Path::new(BUCKET_META_PREFIX)
+        .join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"))
+        .with_extension("rustfs-cas.lock");
+    let lock_path = temp_dir.path().join(RUSTFS_META_BUCKET).join(lock_path);
+
+    manager.cleanup().await.expect("delete checkpoint fixture");
+
+    assert!(
+        !lock_path.exists(),
+        "successful checkpoint cleanup must not leave a task-specific lock artifact"
+    );
+}
+
+#[tokio::test]
+async fn an_empty_blocked_marker_still_blocks_resume_selection() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &blocked_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty blocked marker fixture");
+
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    assert!(CheckpointManager::is_resumable(&disk, &task_id).await.is_err());
+    // Recovery requires replacing/cleaning the snapshot, then removing the
+    // marker; ordinary selector retries are intentionally not an unlock path.
+    manager.cleanup().await.expect("clean blocked checkpoint");
+    assert!(!CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty blocked marker test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_skips_healthy_tasks_with_blocked_markers() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let tasks = [
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::new()),
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::from_static(b"blocked")),
+    ];
+    for (task_id, marker) in &tasks {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.clone(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create healthy resume state");
+        CheckpointManager::new(disk.clone(), task_id.clone())
+            .await
+            .expect("create healthy checkpoint");
+        let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+        let checkpoint_bytes = disk
+            .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read healthy checkpoint before blocking");
+        let marker_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        disk.write_all(RUSTFS_META_BUCKET, &marker_path, marker.clone())
+            .await
+            .expect("write blocked marker");
+
+        assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+                .await
+                .expect("read healthy checkpoint after blocking"),
+            checkpoint_bytes
+        );
+    }
+    temp_dir.close().expect("remove blocked selector test directory");
+}
+
+#[tokio::test]
+async fn stale_checkpoint_manager_cannot_overwrite_newer_progress() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let first = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create first checkpoint manager");
+    let second = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load second checkpoint manager");
+
+    second
+        .update_position(4, 20)
+        .await
+        .expect("persist newer checkpoint progress");
+    let error = first
+        .update_position(1, 3)
+        .await
+        .expect_err("stale checkpoint manager must not overwrite newer progress");
+    assert!(error.to_string().contains("newer progress"));
+
+    let persisted = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load newer checkpoint progress")
+        .get_checkpoint()
+        .await;
+    assert_eq!(persisted.current_bucket_index, 4);
+    assert_eq!(persisted.current_object_index, 20);
+    temp_dir.close().expect("remove stale manager test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_isolates_future_and_corrupt_checkpoints() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let future_task = ResumeUtils::generate_task_id();
+    let corrupt_task = ResumeUtils::generate_task_id();
+    for task_id in [&future_task, &corrupt_task] {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.to_string(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create resumable state fixture");
+    }
+
+    let future_path = format!("{BUCKET_META_PREFIX}/{future_task}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(future_task.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &future_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+    let corrupt_path = format!("{BUCKET_META_PREFIX}/{corrupt_task}_{RESUME_CHECKPOINT_FILE}");
+    let corrupt_bytes = b"{truncated";
+    disk.write_all(RUSTFS_META_BUCKET, &corrupt_path, corrupt_bytes.as_slice().into())
+        .await
+        .expect("write corrupt checkpoint fixture");
+
+    assert!(CheckpointManager::is_resumable(&disk, &future_task).await.is_err());
+    assert!(CheckpointManager::is_resumable(&disk, &corrupt_task).await.is_err());
+    assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
+    for (task_id, path, bytes) in [
+        (&future_task, future_path, future_bytes),
+        (&corrupt_task, corrupt_path, corrupt_bytes.to_vec()),
+    ] {
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &path)
+                .await
+                .expect("read isolated checkpoint bytes"),
+            bytes
+        );
+        let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        assert!(
+            !disk
+                .read_all(RUSTFS_META_BUCKET, &blocked_path)
+                .await
+                .expect("read checkpoint blocked marker")
+                .is_empty()
+        );
+    }
+    temp_dir.close().expect("remove selector isolation test directory");
+}
+
 #[test]
 fn test_persist_throttle_batches_until_threshold() {
     let mut throttle = PersistThrottle::new();

crates/heal/src/heal/resume/utils.rs

@@ -21,7 +21,7 @@ use uuid::Uuid;
 use super::super::{BUCKET_META_PREFIX, DiskError, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
 use super::replacement::{ReplacementPhase, ReplacementRecoveryRecord};
 use super::{
-    EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
+    CheckpointManager, EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
     REPLACEMENT_INTENT_FILE, RESUME_STATE_FILE, ResumeManager, ResumeStateFile, is_replacement_intent, path_to_str,
     replacement_recovery_corruption_for_state_load, replacement_recovery_dir, validate_resume_task_id,
 };
@@ -67,6 +67,7 @@ impl ResumeUtils {
                 // Extract task ID from filename: {task_id}_ahm_resume_state.json
                 if let Some(task_id) = entry.strip_suffix(&format!("_{RESUME_STATE_FILE}"))
                     && validate_resume_task_id(task_id).is_ok()
+                    && CheckpointManager::is_resumable(disk, task_id).await?
                 {
                     task_ids.push(task_id.to_string());
                 }

crates/replication/src/config.rs

@@ -60,7 +60,9 @@ pub const REPLICATION_READ_ONLY_HISTORICAL_FIELDS: &[&str] = &[
     "Destination.ReplicationTime",
 ];
 
-pub const REMOTE_TARGET_CAPABILITY_CONTRACT_VERSION: u32 = 1;
+// v2: disableProxy moved from unsupported to writable (per-target read-proxy
+// opt-out is accepted by set-remote-target and the `proxy` update op).
+pub const REMOTE_TARGET_CAPABILITY_CONTRACT_VERSION: u32 = 2;
 
 pub const REMOTE_TARGET_WRITABLE_FIELDS: &[&str] = &[
     "sourcebucket",
@@ -83,9 +85,12 @@ pub const REMOTE_TARGET_WRITABLE_FIELDS: &[&str] = &[
     // madmin default of 60s); the per-target health-check interval is not
     // yet applied — the heartbeat keeps its global env-configured interval.
     "healthCheckDuration",
+    // Per-target read-proxy opt-out, consumed by the proxy-target selector
+    // (contract v2; previously only importable via MinIO bucket-targets.json).
+    "disableProxy",
 ];
 
-pub const REMOTE_TARGET_UNSUPPORTED_FIELDS: &[&str] = &["disableProxy", "edge", "edgeSyncBeforeExpiry"];
+pub const REMOTE_TARGET_UNSUPPORTED_FIELDS: &[&str] = &["edge", "edgeSyncBeforeExpiry"];
 
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
 pub struct ObjectOpts {

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/data_usage_define/tests.rs

@@ -16,7 +16,7 @@ use super::persistence::DataUsageCacheLoadAttempt;
 use super::*;
 use crate::storage_api::scanner_io::{HTTPRangeSpec, ObjectIO};
 use crate::{ScannerGetObjectReader, ScannerPutObjReader};
-use rustfs_data_usage::{ReplicationAllStats, ReplicationStats};
+use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage};
 use serde_json::Value;
 use std::io::Cursor;
 use std::pin::Pin;
@@ -1636,7 +1636,7 @@ fn size_recursive_prunes_empty_and_preserves_threshold_replication_stats() {
             replication_stats: Some(ReplicationAllStats {
                 targets: HashMap::from([(
                     "arn:test:threshold".to_string(),
-                    ReplicationStats {
+                    ReplicationTargetUsage {
                         after_threshold_count: 1,
                         ..Default::default()
                     },

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]

crates/scanner/src/scanner_io/publish_gate_tests.rs

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 use super::*;
-use rustfs_data_usage::{ReplicationAllStats, ReplicationStats};
+use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage};
 
 const TEST_PLAN_DIGEST: DataUsageScanPlanDigest = DataUsageScanPlanDigest([7; 32]);
 
@@ -271,7 +271,7 @@ fn completed_data_usage_info_flattens_nested_bucket_entries() {
         replication_stats: Some(ReplicationAllStats {
             targets: HashMap::from([(
                 "arn:target".to_string(),
-                ReplicationStats {
+                ReplicationTargetUsage {
                     replicated_size: 2048,
                     replicated_count: 2,
                     ..Default::default()

crates/storage-api/src/lib.rs

@@ -76,6 +76,7 @@ pub use bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOption
 pub use capability::{CapabilitySnapshotError, CapabilityState, CapabilityStatus};
 pub use error::{StorageErrorCode, StorageResult};
 pub use multipart::{CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartUploadResult, PartInfo};
+pub use object::DeleteAccounting;
 pub use object::ObjectLockDeleteOptions;
 pub use object::{DeletedObject, ObjectToDelete};
 pub use object::{ExpirationOptions, TransitionedObject};

crates/storage-api/src/object.rs

@@ -218,6 +218,17 @@ pub struct DeletedObject {
     pub force_delete_generation: Option<i64>,
 }
 
+/// Accounting identity returned by the internal commit-time delete path.
+///
+/// This is carried separately from [`DeletedObject`] so adding quota details
+/// does not change the source shape of the public S3 delete result contract.
+#[derive(Debug, Default, Clone, PartialEq, Eq)]
+pub struct DeleteAccounting {
+    pub size: Option<u64>,
+    pub version_id: Option<Uuid>,
+    pub removed_current_object: bool,
+}
+
 impl DeletedObject {
     pub fn version_purge_status(&self) -> VersionPurgeStatusType {
         self.replication_state
@@ -341,6 +352,19 @@ pub trait ObjectOperations: Send + Sync + fmt::Debug {
         objects: Vec<Self::ObjectToDelete>,
         opts: Self::ObjectOptions,
     ) -> (Vec<Self::DeletedObject>, Vec<Option<Self::Error>>);
+    /// Delete objects and optionally return commit-time accounting identities.
+    /// The default preserves the ordinary delete contract for implementations
+    /// that do not expose storage-level accounting details.
+    async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<Self::ObjectToDelete>,
+        opts: Self::ObjectOptions,
+    ) -> (Vec<Self::DeletedObject>, Vec<Option<Self::Error>>, Vec<Option<DeleteAccounting>>) {
+        let object_count = objects.len();
+        let (deleted, errors) = self.delete_objects(bucket, objects, opts).await;
+        (deleted, errors, vec![None; object_count])
+    }
     async fn put_object_metadata(
         &self,
         bucket: &str,

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/replication.rs

@@ -73,6 +73,8 @@ enum TargetUpdateOp {
     /// Connection group: credentials plus endpoint, target bucket, and TLS settings.
     Credentials,
     Sync,
+    /// Per-target read-proxy opt-out (`disableProxy`).
+    Proxy,
     Bandwidth,
     Path,
 }
@@ -81,12 +83,13 @@ fn parse_remote_target_update_ops(queries: &HashMap<String, String>) -> S3Result
     const SUPPORTED_OPS: &[(&str, TargetUpdateOp)] = &[
         ("creds", TargetUpdateOp::Credentials),
         ("sync", TargetUpdateOp::Sync),
+        ("proxy", TargetUpdateOp::Proxy),
         ("bandwidth", TargetUpdateOp::Bandwidth),
         ("path", TargetUpdateOp::Path),
     ];
     // Present in the MinIO wire contract, but they drive target fields this
     // version rejects as unsupported — fail loudly instead of silently ignoring.
-    const UNSUPPORTED_OPS: &[&str] = &["proxy", "healthcheck", "edge", "edgeSyncBeforeExpiry"];
+    const UNSUPPORTED_OPS: &[&str] = &["healthcheck", "edge", "edgeSyncBeforeExpiry"];
 
     for key in UNSUPPORTED_OPS {
         if queries.get(*key).is_some_and(|value| value == "true") {
@@ -312,11 +315,10 @@ impl RemoteTargetRequest {
             ));
         }
 
-        for (unsupported, configured) in
-            REMOTE_TARGET_UNSUPPORTED_FIELDS
-                .iter()
-                .copied()
-                .zip([self.disable_proxy, self.edge, self.edge_sync_before_expiry])
+        for (unsupported, configured) in REMOTE_TARGET_UNSUPPORTED_FIELDS
+            .iter()
+            .copied()
+            .zip([self.edge, self.edge_sync_before_expiry])
         {
             if configured {
                 return Err(s3_error!(
@@ -702,6 +704,7 @@ impl Operation for SetRemoteTargetHandler {
                         target.deployment_id = remote_target.deployment_id.clone();
                     }
                     TargetUpdateOp::Sync => target.replication_sync = remote_target.replication_sync,
+                    TargetUpdateOp::Proxy => target.disable_proxy = remote_target.disable_proxy,
                     TargetUpdateOp::Bandwidth => target.bandwidth_limit = remote_target.bandwidth_limit,
                     TargetUpdateOp::Path => target.path = remote_target.path.clone(),
                 }
@@ -1520,6 +1523,7 @@ mod tests {
             ("update", "true"),
             ("creds", "true"),
             ("sync", "true"),
+            ("proxy", "true"),
             ("bandwidth", "true"),
             ("path", "true"),
         ]))
@@ -1529,6 +1533,7 @@ mod tests {
             vec![
                 TargetUpdateOp::Credentials,
                 TargetUpdateOp::Sync,
+                TargetUpdateOp::Proxy,
                 TargetUpdateOp::Bandwidth,
                 TargetUpdateOp::Path
             ]
@@ -2070,7 +2075,6 @@ mod tests {
             ("credentials.session_token", serde_json::json!("session-token")),
             ("credentials.expiration", serde_json::json!("2026-01-01T00:00:00Z")),
             ("api", serde_json::json!("s3v2")),
-            ("disableProxy", serde_json::json!(true)),
             ("edge", serde_json::json!(true)),
             ("edgeSyncBeforeExpiry", serde_json::json!(true)),
         ] {
@@ -2300,6 +2304,44 @@ mod tests {
         assert!(!REMOTE_TARGET_UNSUPPORTED_FIELDS.contains(&"healthCheckDuration"));
     }
 
+    #[test]
+    fn remote_target_disable_proxy_is_declared_writable_edge_stays_unsupported() {
+        assert!(REMOTE_TARGET_WRITABLE_FIELDS.contains(&"disableProxy"));
+        assert!(!REMOTE_TARGET_UNSUPPORTED_FIELDS.contains(&"disableProxy"));
+        // edge sync has no implementation behind it — it must stay rejected.
+        assert!(REMOTE_TARGET_UNSUPPORTED_FIELDS.contains(&"edge"));
+        assert!(REMOTE_TARGET_UNSUPPORTED_FIELDS.contains(&"edgeSyncBeforeExpiry"));
+    }
+
+    #[test]
+    fn remote_target_create_accepts_disable_proxy() {
+        let mut request = valid_remote_target_request();
+        request["disableProxy"] = serde_json::json!(true);
+
+        let target = serde_json::from_value::<RemoteTargetRequest>(request)
+            .expect("request should deserialize")
+            .into_bucket_target()
+            .expect("disableProxy is a supported per-target read-proxy opt-out");
+
+        assert!(target.disable_proxy);
+    }
+
+    #[test]
+    fn update_body_with_proxy_op_toggles_disable_proxy_without_credentials() {
+        // Mirrors the other partial-update groups: a proxy-only update body may
+        // omit the connection fields entirely.
+        let body = serde_json::json!({
+            "arn": "arn:rustfs:replication:us-east-1:dep:target",
+            "type": "replication",
+            "disableProxy": true
+        });
+        let request: RemoteTargetRequest = serde_json::from_value(body).expect("partial update body should deserialize");
+        let target = request
+            .into_update_bucket_target(&[TargetUpdateOp::Proxy])
+            .expect("proxy-only update must not require credentials");
+        assert!(target.disable_proxy);
+    }
+
     #[test]
     fn remote_target_capability_fields_do_not_overlap() {
         for field in REMOTE_TARGET_UNSUPPORTED_FIELDS {

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/handlers/system.rs

@@ -1262,7 +1262,9 @@ mod tests {
         assert_eq!(response.summary.manual_transition_jobs.state, CapabilityState::Supported);
         assert_eq!(response.replication.contract_version, 1);
         assert_eq!(response.replication.bucket_replication.contract_version, 1);
-        assert_eq!(response.replication.remote_targets.contract_version, 1);
+        // v2: disableProxy moved from unsupported to writable (per-target
+        // read-proxy opt-out reached the admin API).
+        assert_eq!(response.replication.remote_targets.contract_version, 2);
         assert_eq!(response.replication.bucket_replication.status.state, CapabilityState::Supported);
         assert_eq!(response.replication.remote_targets.status.state, CapabilityState::Supported);
         assert_eq!(
@@ -1293,7 +1295,15 @@ mod tests {
                 .remote_targets
                 .fields
                 .iter()
-                .any(|field| field.name == "disableProxy" && field.state == super::ReplicationFieldState::Unsupported)
+                .any(|field| field.name == "disableProxy" && field.state == super::ReplicationFieldState::Supported)
+        );
+        assert!(
+            response
+                .replication
+                .remote_targets
+                .fields
+                .iter()
+                .any(|field| field.name == "edge" && field.state == super::ReplicationFieldState::Unsupported)
         );
         assert!(
             response
@@ -1364,7 +1374,7 @@ mod tests {
         assert_eq!(value["summary"]["manual_transition_jobs"]["state"], "supported");
         assert_eq!(value["replication"]["contract_version"], 1);
         assert_eq!(value["replication"]["bucket_replication"]["contract_version"], 1);
-        assert_eq!(value["replication"]["remote_targets"]["contract_version"], 1);
+        assert_eq!(value["replication"]["remote_targets"]["contract_version"], 2);
         assert_eq!(value["replication"]["bucket_replication"]["status"]["state"], "supported");
         assert_eq!(value["replication"]["remote_targets"]["status"]["state"], "supported");
         assert_eq!(
@@ -1383,7 +1393,14 @@ mod tests {
                 .as_array()
                 .expect("remote target fields should be an array")
                 .iter()
-                .any(|field| field["name"] == "disableProxy" && field["state"] == "unsupported")
+                .any(|field| field["name"] == "disableProxy" && field["state"] == "supported")
+        );
+        assert!(
+            value["replication"]["remote_targets"]["fields"]
+                .as_array()
+                .expect("remote target fields should be an array")
+                .iter()
+                .any(|field| field["name"] == "edge" && field["state"] == "unsupported")
         );
         assert!(
             value["replication"]["remote_targets"]["fields"]

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!(

rustfs/src/app/bucket_usecase.rs

@@ -1021,7 +1021,7 @@ fn build_list_objects_v2_metadata_output(
                 object: Object {
                     key: Some(encode_list_objects_v2_value(&object.name, encoding_type)),
                     last_modified: object.mod_time.map(Timestamp::from),
-                    size: Some(object.get_actual_size().unwrap_or_default()),
+                    size: Some(object.get_actual_size_or_physical()),
                     e_tag: object.etag.clone().map(|etag| to_s3s_etag(&etag)),
                     storage_class: Some(ObjectStorageClass::from(
                         object

rustfs/src/app/object_usecase.rs

@@ -3969,6 +3969,55 @@ fn delete_creates_delete_marker(opts: &ObjectOptions) -> bool {
     opts.version_id.is_none() && opts.versioned && !opts.version_suspended
 }
 
+fn delete_removes_current_object(opts: &ObjectOptions) -> bool {
+    delete_request_targets_current(
+        opts.version_id
+            .as_deref()
+            .and_then(|version_id| Uuid::parse_str(version_id).ok()),
+    )
+}
+
+fn delete_request_targets_current(version_id: Option<Uuid>) -> bool {
+    version_id.is_none() || version_id.is_some_and(|version_id| version_id.is_nil())
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum DeleteMemoryUpdate {
+    DeleteMarker,
+    Object { size: u64, removed_current_object: bool },
+}
+
+fn delete_memory_update(
+    creates_delete_marker: bool,
+    committed_delete_marker: bool,
+    requested_current: bool,
+    accounting_size: Option<u64>,
+    removed_current_object: bool,
+) -> Option<DeleteMemoryUpdate> {
+    if creates_delete_marker || (committed_delete_marker && requested_current) {
+        return Some(DeleteMemoryUpdate::DeleteMarker);
+    }
+
+    (!committed_delete_marker)
+        .then_some(accounting_size)
+        .flatten()
+        .map(|size| DeleteMemoryUpdate::Object {
+            size,
+            removed_current_object,
+        })
+}
+
+async fn apply_delete_memory_update(bucket: &str, update: Option<DeleteMemoryUpdate>) {
+    match update {
+        Some(DeleteMemoryUpdate::DeleteMarker) => record_bucket_delete_marker_memory(bucket).await,
+        Some(DeleteMemoryUpdate::Object {
+            size,
+            removed_current_object,
+        }) => record_bucket_object_delete_memory(bucket, size, removed_current_object).await,
+        None => {}
+    }
+}
+
 /// `DeleteObjects` is idempotent. A raw filesystem `NotFound` can cross the
 /// distributed delete path instead of its usual typed missing-object error.
 fn is_delete_objects_not_found(error: &EcstoreError) -> bool {
@@ -8409,8 +8458,6 @@ impl DefaultObjectUsecase {
             object: ObjectToDelete,
             versioned: bool,
             version_suspended: bool,
-            size: i64,
-            existing: Option<ObjectInfo>,
         }
 
         // Phase 2 (bounded concurrency, backlog#929 / HP-8): collect the
@@ -8428,32 +8475,23 @@ impl DefaultObjectUsecase {
                     skip_stat,
                 } = prepared;
                 let synthetic_version_id = object.version_id.is_none() && is_dir_object(&object.object_name);
-                let (goi, source_missing) = if skip_stat {
-                    (ObjectInfo::default(), false)
-                } else {
+                if !skip_stat {
                     match store_ref.get_object_info(bucket_ref, &object.object_name, &opts).await {
-                        Ok(res) => (res, false),
-                        Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {
-                            (ObjectInfo::default(), true)
-                        }
+                        Ok(_) => {}
+                        Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
                         Err(err) => return Err(ApiError::from(err)),
                     }
-                };
-
-                let size = goi.size;
+                }
 
                 if synthetic_version_id {
                     object.version_id = Some(Uuid::nil());
                 }
 
-                let existing = (!skip_stat && !source_missing).then_some(goi);
                 Ok::<_, ApiError>(AdmittedDelete {
                     idx,
                     object,
                     versioned: opts.versioned,
                     version_suspended: opts.version_suspended,
-                    size,
-                    existing,
                 })
             }))
             .buffered(DELETE_OBJECTS_PRE_STAT_CONCURRENCY)
@@ -8464,15 +8502,11 @@ impl DefaultObjectUsecase {
         // per-key success/failure reporting is unchanged.
         let mut object_to_delete = Vec::new();
         let mut object_to_delete_idx = Vec::new();
-        let mut object_sizes = Vec::new();
-        let mut existing_object_infos = Vec::new();
         let mut object_versioning = Vec::new();
         for admitted in admitted_deletes {
-            object_sizes.push(admitted.size);
             object_to_delete_idx.push(admitted.idx);
             object_versioning.push((admitted.versioned, admitted.version_suspended));
             object_to_delete.push(admitted.object);
-            existing_object_infos.push(admitted.existing);
         }
         let cache_adapter = self.object_data_cache();
         let cache_keys_before_delete = object_to_delete
@@ -8489,8 +8523,8 @@ impl DefaultObjectUsecase {
             ..Default::default()
         };
         apply_bucket_generation_guard(&req, &bucket, &mut storage_delete_opts)?;
-        let (dobjs, errs) = store
-            .delete_objects_with_tier_delete_journal(&bucket, object_to_delete.clone(), storage_delete_opts)
+        let (dobjs, errs, accounting) = store
+            .delete_objects_with_tier_delete_journal_and_accounting(&bucket, object_to_delete.clone(), storage_delete_opts)
             .await;
 
         let _manager = get_concurrency_manager();
@@ -8515,17 +8549,16 @@ impl DefaultObjectUsecase {
                     delete_results[didx].delete_object = Some(deleted_object.clone());
                     let (versioned, version_suspended) = object_versioning[i];
                     let creates_delete_marker = object_to_delete[i].version_id.is_none() && versioned && !version_suspended;
-                    if creates_delete_marker {
-                        record_bucket_delete_marker_memory(&bucket).await;
-                    } else {
-                        let size = object_sizes[i].max(0) as u64;
-                        record_bucket_object_delete_memory(
-                            &bucket,
-                            size,
-                            existing_object_infos[i].is_some() && object_to_delete[i].version_id.is_none(),
-                        )
-                        .await;
-                    }
+                    let committed_delete_marker = dobjs[i].delete_marker;
+                    let delete_accounting = accounting.get(i).and_then(Option::as_ref);
+                    let update = delete_memory_update(
+                        creates_delete_marker,
+                        committed_delete_marker,
+                        delete_request_targets_current(object_to_delete[i].version_id),
+                        delete_accounting.and_then(|value| value.size),
+                        delete_accounting.is_some_and(|value| value.removed_current_object),
+                    );
+                    apply_delete_memory_update(&bucket, update).await;
                 }
                 Err(error) => {
                     delete_results[didx].error = Some(error);
@@ -8803,12 +8836,24 @@ impl DefaultObjectUsecase {
             let _ = invalidate_object_data_cache_after_delete_success(&cache_adapter, &bucket, &key).await;
         }
 
-        // Fast in-memory update for immediate quota and admin usage consistency
-        if delete_creates_delete_marker(&opts) {
-            record_bucket_delete_marker_memory(&bucket).await;
+        // Fast in-memory update for immediate quota and admin usage consistency.
+        // Prefix/force deletes and synthetic directory entries do not carry one
+        // committed object identity; leave their cache delta to reconciliation.
+        let update = if force_delete || obj_info.name.is_empty() || synthetic_version_id {
+            None
         } else {
-            record_bucket_object_delete_memory(&bucket, obj_info.size.max(0) as u64, opts.version_id.is_none()).await;
-        }
+            // The storage commit returns this object's metadata while its
+            // generation lock is held. Never fall back to a pre-delete stat:
+            // an overwrite can commit between that stat and this delete.
+            delete_memory_update(
+                delete_creates_delete_marker(&opts),
+                obj_info.delete_marker,
+                opts.version_id.is_none(),
+                quota_object_size(&obj_info).ok(),
+                delete_removes_current_object(&opts),
+            )
+        };
+        apply_delete_memory_update(&bucket, update).await;
 
         if obj_info.name.is_empty() {
             if let Some((operation_id, target_arns, generation)) = force_delete_intent {
@@ -17861,6 +17906,158 @@ mod tests {
         assert!(!can_skip_delete_objects_pre_stat(false, &delete_marker_creating_opts(), false));
     }
 
+    #[test]
+    fn delete_accounting_recognizes_explicit_null_as_current_object() {
+        let opts = ObjectOptions {
+            version_id: Some(Uuid::nil().to_string()),
+            version_suspended: true,
+            ..Default::default()
+        };
+        assert!(delete_removes_current_object(&opts));
+        assert!(delete_request_targets_current(Some(Uuid::nil())));
+        assert!(!delete_request_targets_current(Some(Uuid::new_v4())));
+        assert!(!delete_removes_current_object(&ObjectOptions {
+            version_id: Some(Uuid::new_v4().to_string()),
+            ..Default::default()
+        }));
+    }
+
+    #[test]
+    fn compressed_object_delete_restores_usage_baseline() {
+        let mut metadata = HashMap::new();
+        insert_str(&mut metadata, SUFFIX_COMPRESSION, "klauspost/compress/s2".to_string());
+        let object = ObjectInfo {
+            size: 400,
+            actual_size: 1000,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+        let accounting_size = quota_object_size(&object).expect("logical compressed size should be canonical");
+
+        assert_eq!(
+            delete_memory_update(false, false, true, Some(accounting_size), true),
+            Some(DeleteMemoryUpdate::Object {
+                size: 1000,
+                removed_current_object: true,
+            })
+        );
+    }
+
+    #[test]
+    fn invalid_accounting_metadata_is_reconciled_without_overflow() {
+        assert_eq!(delete_memory_update(false, false, true, None, true), None);
+        assert_eq!(
+            delete_memory_update(false, true, true, None, true),
+            Some(DeleteMemoryUpdate::DeleteMarker)
+        );
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn compressed_delete_requests_restore_usage_baseline() {
+        use crate::app::storage_api::test::contract::bucket::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions};
+
+        let store = crate::app::gating_test_env::shared_gating_ecstore().await;
+        if current_app_context().is_none() {
+            crate::app::runtime_sources::install_test_app_context(Arc::clone(&store)).await;
+        }
+        let bucket = format!("compressed-delete-request-{}", Uuid::new_v4().simple());
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create compressed delete request bucket");
+
+        // Seed the process-local usage with the canonical logical bytes. The
+        // direct storage PUT below intentionally does not apply an app-layer
+        // usage delta; the two real DELETE requests must remove exactly this
+        // amount through their request-layer wiring.
+        crate::app::storage_api::test::data_usage::seed_bucket_usage_memory_for_test(&bucket, 2_000).await;
+
+        for object in ["single", "batch"] {
+            let mut metadata = HashMap::new();
+            insert_str(&mut metadata, SUFFIX_COMPRESSION, "klauspost/compress/s2".to_string());
+            insert_str(&mut metadata, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+            let reader = HashReader::from_stream(std::io::Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid");
+            let mut reader = PutObjReader::new(reader);
+            store
+                .put_object(
+                    &bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        user_defined: metadata,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("compressed fixture object should be written");
+        }
+
+        let mut single_req = build_request(
+            DeleteObjectInput::builder()
+                .bucket(bucket.clone())
+                .key("single".to_string())
+                .build()
+                .expect("single delete input should build"),
+            Method::DELETE,
+        );
+        single_req.extensions.insert(crate::storage::access::ReqInfo {
+            cred: Some(rustfs_credentials::Credentials::default()),
+            is_owner: true,
+            ..Default::default()
+        });
+        DefaultObjectUsecase::from_global()
+            .execute_delete_object(single_req)
+            .await
+            .expect("single compressed delete should succeed");
+        assert_eq!(
+            crate::app::storage_api::test::data_usage::get_bucket_usage_memory(&bucket).await,
+            Some(1_000),
+            "single delete must subtract the logical accounting size"
+        );
+
+        let mut batch_req = build_request(
+            DeleteObjectsInput::builder()
+                .bucket(bucket.clone())
+                .delete(Delete {
+                    objects: vec![ObjectIdentifier {
+                        key: "batch".to_string(),
+                        ..Default::default()
+                    }],
+                    quiet: None,
+                })
+                .build()
+                .expect("batch delete input should build"),
+            Method::POST,
+        );
+        batch_req.extensions.insert(crate::storage::access::ReqInfo {
+            cred: Some(rustfs_credentials::Credentials::default()),
+            is_owner: true,
+            ..Default::default()
+        });
+        DefaultObjectUsecase::from_global()
+            .execute_delete_objects(batch_req)
+            .await
+            .expect("batch compressed delete should succeed");
+        assert_eq!(
+            crate::app::storage_api::test::data_usage::get_bucket_usage_memory(&bucket).await,
+            Some(0),
+            "batch delete must subtract the committed logical accounting size"
+        );
+
+        store
+            .delete_bucket(
+                &bucket,
+                &DeleteBucketOptions {
+                    force: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("clean up compressed delete request bucket");
+    }
+
     #[tokio::test]
     async fn execute_get_object_attributes_returns_internal_error_when_store_uninitialized() {
         let input = GetObjectAttributesInput::builder()

rustfs/src/app/storage_api.rs

@@ -72,6 +72,11 @@ pub(crate) mod data_usage {
         compute_bucket_usage, live_bucket_usage_computations, seed_bucket_usage_memory_for_test, store_data_usage_in_backend,
     };
 
+    #[cfg(test)]
+    pub(crate) async fn get_bucket_usage_memory(bucket: &str) -> Option<u64> {
+        crate::storage::storage_api::ecstore_data_usage::get_bucket_usage_memory(bucket).await
+    }
+
     pub(crate) async fn record_bucket_object_delete_memory(bucket: &str, deleted_size: u64, removed_current_object: bool) {
         crate::storage::storage_api::ecstore_data_usage::record_bucket_object_delete_memory(
             bucket,
@@ -1233,7 +1238,10 @@ pub(crate) mod test {
 
     pub(crate) use super::access::ReqInfo;
     pub(crate) use super::options::VERSIONING_CONFIG_LOOKUPS;
-    pub(crate) use super::{bucket, data_usage, ecfs, object_utils, runtime};
+    pub(crate) use super::{bucket, ecfs, object_utils, runtime};
+    pub(crate) mod data_usage {
+        pub(crate) use super::super::data_usage::*;
+    }
     pub(crate) use crate::storage::storage_api::test_consumer::{get_global_bucket_metadata_sys, set_bucket_metadata};
     pub(crate) use crate::storage::storage_api::{
         ECStore, Endpoint, Endpoints, PoolEndpoints, StorageObjectInfo, StorageObjectOptions, StoragePutObjReader,

rustfs/src/storage/s3_api/bucket.rs

@@ -296,7 +296,10 @@ pub(crate) fn build_list_objects_v2_output(
             let mut obj = Object {
                 key: Some(key),
                 last_modified: v.mod_time.map(Timestamp::from),
-                size: Some(v.get_actual_size().unwrap_or_default()),
+                // Compressed legacy objects may retain an unknown (-1)
+                // logical-size sentinel; never expose that internal value in
+                // an S3 response.
+                size: Some(v.get_actual_size_or_physical()),
                 e_tag: v.etag.clone().map(|etag| to_s3s_etag(&etag)),
                 storage_class: v.storage_class.clone().map(ObjectStorageClass::from),
                 ..Default::default()
@@ -656,6 +659,45 @@ mod tests {
         assert_eq!(output.common_prefixes.as_ref().map(std::vec::Vec::len), Some(2));
     }
 
+    #[test]
+    fn list_objects_never_exposes_compressed_unknown_size_sentinel() {
+        let mut metadata = std::collections::HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let output = build_list_objects_v2_output(
+            ListObjectsV2Info {
+                objects: vec![ObjectInfo {
+                    name: "legacy-compressed".to_string(),
+                    size: 128,
+                    actual_size: -1,
+                    user_defined: std::sync::Arc::new(metadata),
+                    ..Default::default()
+                }],
+                ..Default::default()
+            },
+            false,
+            1000,
+            "bucket".to_string(),
+            String::new(),
+            None,
+            None,
+            None,
+            None,
+        );
+
+        assert_eq!(
+            output
+                .contents
+                .as_ref()
+                .and_then(|objects| objects.first())
+                .and_then(|object| object.size),
+            Some(128)
+        );
+    }
+
     #[test]
     fn list_responses_report_standard_for_legacy_label_only_file_metadata() {
         let version_id = Uuid::parse_str("11111111-2222-3333-4444-555555555555").expect("fixture version ID should be valid");

rustfs/src/storage/storage_api.rs

@@ -429,6 +429,8 @@ pub(crate) mod ecstore_config {
 }
 
 pub(crate) mod ecstore_data_usage {
+    #[cfg(test)]
+    pub(crate) use rustfs_ecstore::api::data_usage::get_bucket_usage_memory;
     pub(crate) use rustfs_ecstore::api::data_usage::{
         apply_bucket_usage_memory_overlay, init_compression_total_memory_from_backend, load_admin_data_usage_from_backend_cached,
         load_data_usage_from_backend, quota_object_size, record_bucket_delete_marker_memory, record_bucket_object_delete_memory,