fix(heal): preserve resumable bucket checkpoints

* test(ecstore): cover suspended-owner heal semantics

* test(heal): cover suspended owner production path

.github/workflows/nightly-gnu.yml

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

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/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/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;
@@ -5655,7 +5656,18 @@ 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 +5757,7 @@ 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 +5804,22 @@ 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 +5948,12 @@ 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 +5999,7 @@ 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 +6237,16 @@ 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 +6575,12 @@ 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 +7872,113 @@ 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/heal/src/heal/erasure_healer.rs

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 use crate::heal::{
-    progress::HealProgress,
+    progress::{HealProgress, add_bytes, increment_counter},
     resume::{
-        CheckpointManager, ReplacementTargetIdentity, ResumeManager, ResumeUtils, compose_key,
+        CheckpointManager, CheckpointObjectOutcome, ReplacementTargetIdentity, ResumeManager, ResumeUtils, compose_key,
         replacement_target_identities_match,
     },
     storage::{HealStorageAPI, next_heal_listing_token},
@@ -410,6 +410,9 @@ impl ErasureSetHealer {
                 && state.successful_objects == 0
                 && state.failed_objects == 0
                 && state.skipped_objects == 0
+                && state.skipped_new_versions == 0
+                && state.skipped_ilm_expired == 0
+                && state.processed_bytes == 0
             {
                 // schedule_retry persists the authoritative resume reset before
                 // resetting the checkpoint. Reapply the checkpoint reset after
@@ -474,6 +477,23 @@ impl ErasureSetHealer {
 
         // 2. initialize progress
         self.initialize_progress(buckets, &state).await;
+        let (baseline_known, baseline_count, baseline_size, baseline_generation) = {
+            let baseline = self.progress.read().await;
+            (
+                baseline.baseline_known,
+                baseline.objects_total_count,
+                baseline.objects_total_size,
+                baseline.baseline_generation,
+            )
+        };
+        if baseline_known {
+            resume_manager
+                .set_progress_baseline(baseline_count, baseline_size, baseline_generation)
+                .await?;
+            checkpoint_manager
+                .set_progress_baseline(baseline_count, baseline_size, baseline_generation)
+                .await?;
+        }
 
         // 3. continue from checkpoint
         let current_bucket_index = checkpoint.current_bucket_index;
@@ -483,12 +503,66 @@ impl ErasureSetHealer {
         let mut successful_objects = state.successful_objects;
         let mut failed_objects = state.failed_objects;
         let mut skipped_objects = state.skipped_objects;
+        let checkpoint_has_progress = checkpoint.baseline_known
+            || checkpoint.successful_objects > 0
+            || checkpoint.failed_object_count > 0
+            || checkpoint.skipped_object_count > 0
+            || checkpoint.skipped_new_versions > 0
+            || checkpoint.skipped_ilm_expired > 0
+            || checkpoint.processed_bytes > 0
+            || checkpoint.total_objects > 0
+            || checkpoint.total_bytes > 0
+            || checkpoint.baseline_generation.is_some()
+            || checkpoint.counter_unknown;
+        let checkpoint_generation_mismatch = checkpoint.baseline_known && checkpoint.baseline_generation != baseline_generation;
+        let mut restored_counter_unknown = state.counter_unknown || checkpoint.counter_unknown;
+        if checkpoint_has_progress {
+            successful_objects = checkpoint.successful_objects;
+            failed_objects = checkpoint.failed_object_count;
+            skipped_objects = checkpoint.skipped_object_count;
+            let restored_processed_objects = successful_objects
+                .checked_add(failed_objects)
+                .and_then(|value| value.checked_add(skipped_objects))
+                .and_then(|value| value.checked_add(checkpoint.skipped_new_versions))
+                .and_then(|value| value.checked_add(checkpoint.skipped_ilm_expired));
+            let checkpoint_counter_overflow = restored_processed_objects.is_none();
+            restored_counter_unknown |= checkpoint_counter_overflow;
+            processed_objects = restored_processed_objects.unwrap_or(u64::MAX);
+            let mut progress = self.progress.write().await;
+            progress.objects_scanned = processed_objects;
+            progress.objects_healed = successful_objects;
+            progress.objects_failed = failed_objects;
+            progress.skipped_objects = skipped_objects;
+            progress.skipped_new_versions = checkpoint.skipped_new_versions;
+            progress.skipped_ilm_expired = checkpoint.skipped_ilm_expired;
+            if checkpoint.baseline_known && !checkpoint_generation_mismatch {
+                progress.objects_total_count = checkpoint.total_objects;
+                progress.objects_total_size = checkpoint.total_bytes;
+                progress.baseline_generation = checkpoint.baseline_generation;
+                progress.baseline_known = true;
+            }
+            progress.bytes_processed = checkpoint.processed_bytes;
+            progress.counter_unknown = state.counter_unknown || checkpoint.counter_unknown;
+            progress.refresh_progress_percentage();
+            if checkpoint_generation_mismatch || checkpoint_counter_overflow || progress.counter_unknown {
+                progress.mark_unknown();
+            }
+        }
+        if checkpoint_generation_mismatch {
+            restored_counter_unknown = true;
+        }
+        if restored_counter_unknown {
+            checkpoint_manager.mark_counter_unknown().await?;
+            resume_manager.mark_counter_unknown().await?;
+        }
         let mut failed_buckets = 0u64;
 
         // 4. process remaining buckets
         for (bucket_idx, bucket) in buckets.iter().enumerate().skip(current_bucket_index) {
             // check if completed
             if state.completed_buckets.contains(bucket) {
+                checkpoint_manager.complete_bucket(bucket_idx.saturating_add(1)).await?;
+                current_object_index = 0;
                 continue;
             }
 
@@ -516,13 +590,42 @@ impl ErasureSetHealer {
                 return bucket_result;
             }
 
-            // update checkpoint position
-            checkpoint_manager.update_position(bucket_idx, current_object_index).await?;
-
             // update progress
-            resume_manager
-                .update_progress(processed_objects, successful_objects, failed_objects, skipped_objects)
+            let progress_snapshot = self.progress.read().await;
+            let bytes_processed = progress_snapshot.bytes_processed;
+            let skipped_new_versions = progress_snapshot.skipped_new_versions;
+            let skipped_ilm_expired = progress_snapshot.skipped_ilm_expired;
+            let counter_unknown = progress_snapshot.counter_unknown;
+            drop(progress_snapshot);
+            // The checkpoint is the recovery authority for object progress.
+            // Publish its counters and fence before the resume summary so a
+            // crash between the two stores cannot make recovery select newer
+            // summary bytes with an older checkpoint ledger.
+            if counter_unknown {
+                checkpoint_manager.mark_counter_unknown().await?;
+            }
+            checkpoint_manager
+                .update_progress(successful_objects, failed_objects, skipped_objects, bytes_processed)
                 .await?;
+            checkpoint_manager
+                .set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired)
+                .await?;
+            checkpoint_manager.update_position(bucket_idx, current_object_index).await?;
+            resume_manager
+                .update_progress_with_bytes(
+                    processed_objects,
+                    successful_objects,
+                    failed_objects,
+                    skipped_objects,
+                    bytes_processed,
+                )
+                .await?;
+            resume_manager
+                .set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired)
+                .await?;
+            if counter_unknown {
+                resume_manager.mark_counter_unknown().await?;
+            }
 
             // check cancel status
             if self.cancel_token.is_cancelled() {
@@ -542,6 +645,7 @@ impl ErasureSetHealer {
             match bucket_result {
                 Ok(_) => {
                     resume_manager.complete_bucket(bucket).await?;
+                    checkpoint_manager.complete_bucket(bucket_idx.saturating_add(1)).await?;
                     debug!(
                         target: "rustfs::heal::erasure_healer",
                         event = EVENT_HEAL_ERASURE_BUCKET_STATE,
@@ -567,7 +671,9 @@ impl ErasureSetHealer {
                         error = %e,
                         "Erasure set bucket heal failed"
                     );
-                    // continue to next bucket, do not interrupt the whole process
+                    // A single durable cursor and ledger cannot safely preserve
+                    // this bucket while processing a later one.
+                    break;
                 }
             }
 
@@ -775,20 +881,48 @@ impl ErasureSetHealer {
 
                 // Per-version dedup identity — the single canonical key.
                 let key = compose_key(&item.name, item.version_id.as_deref());
-                if checkpoint.processed_objects.contains(&key) || checkpoint.skipped_objects.contains(&key) {
+                if checkpoint.processed_objects.contains(&key)
+                    || checkpoint.failed_objects.contains(&key)
+                    || checkpoint.skipped_objects.contains(&key)
+                {
                     continue;
                 }
 
                 if should_skip_new_version(item.mod_time_unix_nanos, started_at_secs) {
-                    checkpoint_manager.add_processed_object(key).await?;
-                    *processed_objects = processed_objects.saturating_add(1);
+                    let counter_ok = increment_counter(processed_objects);
                     completed_in_page = completed_in_page.saturating_add(1);
                     counter!("rustfs_heal_skipped_new_versions_total").increment(1);
-                    {
+                    let (skipped_new, skipped_ilm, counter_unknown) = {
                         let mut progress = self.progress.write().await;
                         progress.record_skipped_new_version();
                         progress.set_current_object(Some(format!("skipped_new: {bucket}/{}", item.name)));
-                        progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                        progress.update_object_progress(
+                            *processed_objects,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                        );
+                        if !counter_ok {
+                            progress.mark_unknown();
+                        }
+                        (progress.skipped_new_versions, progress.skipped_ilm_expired, progress.counter_unknown)
+                    };
+                    checkpoint_manager
+                        .record_object_outcome(
+                            key,
+                            CheckpointObjectOutcome::Processed,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                            skipped_new,
+                            skipped_ilm,
+                            !counter_ok || counter_unknown,
+                        )
+                        .await?;
+                    if !counter_ok || counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                     }
                     debug!(
                         target: "rustfs::heal::erasure_healer",
@@ -820,15 +954,40 @@ impl ErasureSetHealer {
                         )
                         .await?
                 {
-                    checkpoint_manager.add_processed_object(key).await?;
-                    *processed_objects = processed_objects.saturating_add(1);
+                    let counter_ok = increment_counter(processed_objects);
                     completed_in_page = completed_in_page.saturating_add(1);
                     counter!("rustfs_heal_skipped_ilm_expired_total").increment(1);
-                    {
+                    let (skipped_new, skipped_ilm, counter_unknown) = {
                         let mut progress = self.progress.write().await;
                         progress.record_skipped_ilm_expired();
                         progress.set_current_object(Some(format!("skipped_ilm: {bucket}/{}", item.name)));
-                        progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                        progress.update_object_progress(
+                            *processed_objects,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                        );
+                        if !counter_ok {
+                            progress.mark_unknown();
+                        }
+                        (progress.skipped_new_versions, progress.skipped_ilm_expired, progress.counter_unknown)
+                    };
+                    checkpoint_manager
+                        .record_object_outcome(
+                            key,
+                            CheckpointObjectOutcome::Processed,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                            skipped_new,
+                            skipped_ilm,
+                            !counter_ok || counter_unknown,
+                        )
+                        .await?;
+                    if !counter_ok || counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                     }
                     debug!(
                         target: "rustfs::heal::erasure_healer",
@@ -954,11 +1113,11 @@ impl ErasureSetHealer {
 
             while let Some((key, object, version_id, result)) = page_tasks.next().await {
                 let (object_size, result) = result;
-                match result {
+                let mut telemetry_unknown = false;
+                let checkpoint_outcome = match result {
                     Ok(true) => {
-                        *successful_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_processed_object(key).await?;
+                        telemetry_unknown |= !increment_counter(successful_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                         debug!(
                             target: "rustfs::heal::erasure_healer",
                             event = EVENT_HEAL_ERASURE_OBJECT_STATE,
@@ -971,11 +1130,11 @@ impl ErasureSetHealer {
                             state = "healed",
                             "Erasure set object healed"
                         );
+                        CheckpointObjectOutcome::Processed
                     }
                     Ok(false) => {
-                        checkpoint_manager.add_processed_object(key).await?;
-                        *successful_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
+                        telemetry_unknown |= !increment_counter(successful_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                         debug!(
                             target: "rustfs::heal::erasure_healer",
                             event = EVENT_HEAL_ERASURE_OBJECT_STATE,
@@ -988,12 +1147,12 @@ impl ErasureSetHealer {
                             state = "missing_treated_as_ok",
                             "Erasure set missing object treated as ok"
                         );
+                        CheckpointObjectOutcome::Processed
                     }
                     Err(err @ Error::TaskCancelled) | Err(err @ Error::TaskTimeout) => return Err(err),
                     Err(Error::TransientSkip { message }) => {
-                        *skipped_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_skipped_object(key).await?;
+                        telemetry_unknown |= !increment_counter(skipped_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                         demote_to_debug_when!(!take_failure_log_sample(&mut transient_skip_samples_logged), warn, target: "rustfs::heal::erasure_healer", {
                             event = EVENT_HEAL_ERASURE_OBJECT_STATE,
                             component = LOG_COMPONENT_HEAL,
@@ -1006,11 +1165,11 @@ impl ErasureSetHealer {
                             error = %message,
                             "Erasure set object heal skipped due to transient error"
                         });
+                        CheckpointObjectOutcome::Skipped
                     }
                     Err(err) => {
-                        *failed_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_failed_object(key).await?;
+                        telemetry_unknown |= !increment_counter(failed_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                         demote_to_debug_when!(!take_failure_log_sample(&mut failure_samples_logged), warn, target: "rustfs::heal::erasure_healer", {
                             event = EVENT_HEAL_ERASURE_OBJECT_STATE,
                             component = LOG_COMPONENT_HEAL,
@@ -1023,15 +1182,42 @@ impl ErasureSetHealer {
                             error = %err,
                             "Erasure set object heal failed"
                         });
+                        CheckpointObjectOutcome::Failed
                     }
-                }
+                };
 
-                *processed_objects += 1;
+                telemetry_unknown |= !increment_counter(processed_objects);
                 completed_in_page += 1;
-                {
+                let (progress_unknown, skipped_new_versions, skipped_ilm_expired) = {
                     let mut progress = self.progress.write().await;
                     progress.set_current_object(Some(format!("{bucket}/{object}")));
-                    progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                    progress.update_object_progress(
+                        *processed_objects,
+                        *successful_objects,
+                        *failed_objects,
+                        *skipped_objects,
+                        bytes_processed,
+                    );
+                    if telemetry_unknown {
+                        progress.mark_unknown();
+                    }
+                    (progress.counter_unknown, progress.skipped_new_versions, progress.skipped_ilm_expired)
+                };
+                checkpoint_manager
+                    .record_object_outcome(
+                        key,
+                        checkpoint_outcome,
+                        *successful_objects,
+                        *failed_objects,
+                        *skipped_objects,
+                        bytes_processed,
+                        skipped_new_versions,
+                        skipped_ilm_expired,
+                        telemetry_unknown || progress_unknown,
+                    )
+                    .await?;
+                if telemetry_unknown || progress_unknown {
+                    resume_manager.mark_counter_unknown().await?;
                 }
 
                 if completed_in_page.is_multiple_of(100) {
@@ -1041,16 +1227,22 @@ impl ErasureSetHealer {
 
             *current_object_index = global_obj_idx;
 
-            // Persist the authoritative cursor FIRST (points at the next page
-            // boundary), then prune the per-version dedup sets. Both are
-            // idempotent under crash: heal_object re-heals safely.
-            let next_cursor = if is_truncated { next_token.clone() } else { None };
-            resume_manager.set_resume_cursor(next_cursor.clone()).await?;
-            checkpoint_manager.complete_page(bucket_index, *current_object_index).await?;
+            // Persist the checkpoint ledger and page position before exposing
+            // the next resume cursor. A crash before cursor publication keeps
+            // the page identities available for exact-once replay.
+            checkpoint_manager.advance_page(bucket_index, *current_object_index).await?;
             // Check if there are more pages
             if !is_truncated {
                 break;
             }
+            continuation_token = next_heal_listing_token(bucket, "", next_token, is_truncated)?;
+            if continuation_token.is_none() {
+                // A truncated page without a continuation token is terminal.
+                // Retain its ledger until bucket completion is durable.
+                break;
+            }
+            resume_manager.set_resume_cursor(continuation_token.clone()).await?;
+            checkpoint_manager.prune_completed_page().await?;
 
             // Anti-loop guard: an empty page reported as truncated cannot advance
             // the cursor (there is no last identity to move past), so treat it as a
@@ -1069,12 +1261,6 @@ impl ErasureSetHealer {
                 )));
             }
             previous_page_last = page_last;
-
-            continuation_token = next_heal_listing_token(bucket, "", next_token, is_truncated)?;
-            if continuation_token.is_none() {
-                // Truncated but no continuation token: treat as end of listing.
-                break;
-            }
         }
 
         Ok(())
@@ -1083,10 +1269,66 @@ impl ErasureSetHealer {
     /// initialize progress tracking
     async fn initialize_progress(&self, _buckets: &[String], state: &crate::heal::resume::ResumeState) {
         let mut progress = self.progress.write().await;
-        progress.objects_scanned = state.total_objects;
+        let existing_baseline = (
+            progress.objects_total_count,
+            progress.objects_total_size,
+            progress.baseline_generation,
+            progress.progress_state,
+            progress.baseline_known,
+        );
+        let baseline_generation_mismatch =
+            state.baseline_known && existing_baseline.4 && state.baseline_generation != existing_baseline.2;
+        let use_persisted_baseline = state.baseline_known && !baseline_generation_mismatch;
+        progress.objects_scanned = state.processed_objects;
         progress.objects_healed = state.successful_objects;
         progress.objects_failed = state.failed_objects;
-        progress.bytes_processed = 0; // Resume state tracks object counts, not byte counters.
+        progress.skipped_objects = state.skipped_objects;
+        progress.skipped_new_versions = state.skipped_new_versions;
+        progress.skipped_ilm_expired = state.skipped_ilm_expired;
+        progress.bytes_processed = state.processed_bytes;
+        progress.counter_unknown = state.counter_unknown;
+        if use_persisted_baseline
+            || existing_baseline.0 > 0
+            || existing_baseline.1 > 0
+            || existing_baseline.2.is_some()
+            || existing_baseline.4
+        {
+            progress.objects_total_count = if use_persisted_baseline {
+                state.total_objects
+            } else {
+                existing_baseline.0
+            };
+            progress.objects_total_size = if use_persisted_baseline {
+                state.total_bytes
+            } else {
+                existing_baseline.1
+            };
+            progress.baseline_generation = if use_persisted_baseline {
+                state.baseline_generation
+            } else {
+                existing_baseline.2
+            };
+            progress.baseline_known = use_persisted_baseline
+                || existing_baseline.0 > 0
+                || existing_baseline.1 > 0
+                || existing_baseline.2.is_some()
+                || existing_baseline.4;
+        }
+        progress.progress_state = if use_persisted_baseline
+            || existing_baseline.0 > 0
+            || existing_baseline.1 > 0
+            || existing_baseline.2.is_some()
+            || existing_baseline.4
+        {
+            crate::heal::progress::HealProgressState::Running
+        } else {
+            crate::heal::progress::HealProgressState::Indeterminate
+        };
+        if baseline_generation_mismatch || state.counter_unknown {
+            progress.mark_unknown();
+        }
+        progress.ledger_complete = false;
+        progress.refresh_progress_percentage();
         progress.start_time = UNIX_EPOCH.checked_add(Duration::from_secs(state.start_time));
         progress.last_update_time = UNIX_EPOCH.checked_add(Duration::from_secs(state.last_update));
         progress.set_current_object(state.current_object.clone());
@@ -1264,8 +1506,8 @@ mod resume_loop_tests {
     };
     use crate::heal::progress::HealProgress;
     use crate::heal::resume::{
-        CheckpointManager, RESUME_CHECKPOINT_FILE, ReplacementTargetIdentity, ResumeDeleteFailure, ResumeManager, ResumeUtils,
-        compose_key,
+        CheckpointManager, CheckpointObjectOutcome, RESUME_CHECKPOINT_FILE, ReplacementTargetIdentity, ResumeDeleteFailure,
+        ResumeManager, ResumeUtils, compose_key,
     };
     use crate::heal::storage::{HealLifecycleExpiryContext, HealListItem, HealObjectInfo, HealStorageAPI};
     use crate::heal::storage_api::status::BucketInfo;
@@ -1405,6 +1647,7 @@ mod resume_loop_tests {
         list_include_lifecycle_object_info: Mutex<Vec<bool>>,
         replacement_target_identity_sequences: Mutex<VecDeque<Vec<ReplacementTargetIdentity>>>,
         fail_listing: AtomicBool,
+        fail_listing_buckets: Mutex<HashSet<String>>,
     }
 
     impl FakeStorage {
@@ -1441,6 +1684,9 @@ mod resume_loop_tests {
         fn fail_listing(&self) {
             self.fail_listing.store(true, Ordering::SeqCst);
         }
+        fn fail_bucket_listing(&self, bucket: &str) {
+            self.fail_listing_buckets.lock().unwrap().insert(bucket.to_string());
+        }
     }
 
     #[async_trait::async_trait]
@@ -1531,7 +1777,7 @@ mod resume_loop_tests {
         }
         async fn list_objects_for_heal_page(
             &self,
-            _bucket: &str,
+            bucket: &str,
             _prefix: &str,
             continuation_token: Option<&str>,
             include_lifecycle_object_info: bool,
@@ -1540,7 +1786,7 @@ mod resume_loop_tests {
                 .lock()
                 .unwrap()
                 .push(include_lifecycle_object_info);
-            if self.fail_listing.load(Ordering::SeqCst) {
+            if self.fail_listing.load(Ordering::SeqCst) || self.fail_listing_buckets.lock().unwrap().contains(bucket) {
                 return Err(Error::other("injected listing failure"));
             }
             let key = continuation_token.map(str::to_string);
@@ -1918,6 +2164,49 @@ mod resume_loop_tests {
         assert!(state.completed_buckets.is_empty(), "the failed bucket must remain resumable");
     }
 
+    #[tokio::test]
+    async fn bucket_failure_stops_before_a_later_bucket_checkpoint() {
+        let env = make_env().await;
+        let task_id = ResumeUtils::generate_task_id();
+        let buckets = vec!["a".to_string(), "b".to_string()];
+        let resume = ResumeManager::new(
+            env.healer.disk.clone(),
+            task_id.clone(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            buckets.clone(),
+        )
+        .await
+        .unwrap();
+        let checkpoint = CheckpointManager::new(env.healer.disk.clone(), task_id.clone())
+            .await
+            .unwrap();
+        env.storage.fail_bucket_listing("a");
+        for _ in 0..3 {
+            assert!(resume.schedule_retry().await.unwrap());
+        }
+
+        env.healer
+            .execute_heal_with_resume(&buckets, "pool_0_set_0", &resume, &checkpoint)
+            .await
+            .expect_err("the first bucket failure must keep the pass incomplete");
+        let persisted = checkpoint.get_checkpoint().await;
+        assert_eq!(persisted.current_bucket_index, 0);
+        assert!(resume.get_state().await.completed_buckets.is_empty());
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&buckets, "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect_err("recovery must retry the earlier failed bucket");
+        assert!(!resumed.get_state().await.completed);
+    }
+
     #[tokio::test]
     async fn completed_resume_state_is_not_selected_for_a_new_heal() {
         let env = make_env().await;
@@ -2107,8 +2396,175 @@ mod resume_loop_tests {
         let mut names: Vec<String> = env.storage.calls().into_iter().map(|(n, _)| n).collect();
         names.sort();
         assert_eq!(names, vec!["a", "b", "c", "d"], "every object exactly once, none dropped/doubled");
-        // Final page not truncated => cursor cleared.
-        assert_eq!(env.resume.resume_cursor().await, None);
+        // Keep the final page cursor until the outer loop durably completes the
+        // bucket, so a crash can replay only this page against its identities.
+        assert_eq!(env.resume.resume_cursor().await, Some("t1".to_string()));
+    }
+
+    #[tokio::test]
+    async fn persisted_failure_waits_for_the_bounded_retry_after_page_replay() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+        env.checkpoint
+            .record_object_outcome(
+                compose_key("object", Some("v1")),
+                CheckpointObjectOutcome::Failed,
+                0,
+                1,
+                0,
+                0,
+                0,
+                0,
+                false,
+            )
+            .await
+            .unwrap();
+        env.checkpoint.advance_page(0, 1).await.unwrap();
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect_err("the persisted failure must schedule a bounded retry");
+        assert!(
+            env.storage.calls().is_empty(),
+            "the failed identity must not be repeated in the same pass"
+        );
+
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("the bounded retry must heal the object");
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 1);
+        assert_eq!(state.failed_objects, 0);
+    }
+
+    #[tokio::test]
+    async fn final_page_crash_replays_only_the_retained_page_identities() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("first", Some("v1"), false)],
+                next: Some("final-page".to_string()),
+                truncated: true,
+            },
+        );
+        env.storage.set_page(
+            Some("final-page"),
+            Page {
+                items: vec![item("last", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+
+        let (processed, successful, failed, skipped, result) = run(&env).await;
+        result.expect("the bucket pass must finish before the simulated crash");
+        assert_eq!((processed, successful, failed, skipped), (2, 2, 0, 0));
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("the retained final-page ledger must make recovery exact");
+
+        assert_eq!(
+            env.storage.calls(),
+            vec![
+                ("first".to_string(), Some("v1".to_string())),
+                ("last".to_string(), Some("v1".to_string()))
+            ]
+        );
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 2);
+        assert_eq!(state.processed_objects, 2);
+    }
+
+    #[tokio::test]
+    async fn truncated_page_without_token_retains_its_replay_ledger() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: true,
+            },
+        );
+
+        let (processed, successful, failed, skipped, result) = run(&env).await;
+        result.expect("the tokenless truncated page is a terminal page");
+        assert_eq!((processed, successful, failed, skipped), (1, 1, 0, 0));
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("terminal-page recovery must not replay a durable identity");
+
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 1);
+        assert_eq!(state.processed_objects, 1);
+    }
+
+    #[tokio::test]
+    async fn completed_bucket_reconciles_its_final_page_checkpoint_after_crash() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+
+        let (_, _, _, _, result) = run(&env).await;
+        result.expect("the bucket pass must finish before the simulated crash");
+        env.resume.complete_bucket("b").await.unwrap();
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("recovery must finish the checkpoint transition without replaying the bucket");
+
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let checkpoint = checkpoint.get_checkpoint().await;
+        assert_eq!(checkpoint.current_bucket_index, 1);
+        assert!(checkpoint.processed_objects.is_empty());
     }
 
     #[tokio::test]

crates/heal/src/heal/manager.rs

@@ -2012,16 +2012,44 @@ impl HealManager {
         }
 
         let mut snapshot = HealProgress::default();
+        let mut has_object_sweep = false;
+        let mut all_object_baselines_known = true;
+        let mut counter_overflow = false;
+        let mut stage_current = 0_u64;
+        let mut stage_total = 0_u64;
         for task in active_tasks {
             let progress = task.get_progress().await;
-            snapshot.objects_scanned = snapshot.objects_scanned.saturating_add(progress.objects_scanned);
-            snapshot.objects_healed = snapshot.objects_healed.saturating_add(progress.objects_healed);
-            snapshot.objects_failed = snapshot.objects_failed.saturating_add(progress.objects_failed);
-            snapshot.skipped_new_versions = snapshot.skipped_new_versions.saturating_add(progress.skipped_new_versions);
-            snapshot.skipped_ilm_expired = snapshot.skipped_ilm_expired.saturating_add(progress.skipped_ilm_expired);
-            snapshot.objects_total_count = snapshot.objects_total_count.saturating_add(progress.objects_total_count);
-            snapshot.objects_total_size = snapshot.objects_total_size.saturating_add(progress.objects_total_size);
-            snapshot.bytes_processed = snapshot.bytes_processed.saturating_add(progress.bytes_processed);
+            let object_sweep = matches!(progress.kind, crate::heal::progress::HealProgressKind::ObjectSweep);
+            has_object_sweep |= object_sweep;
+            if object_sweep {
+                all_object_baselines_known &= progress.baseline_known;
+            }
+            counter_overflow |=
+                progress.counter_unknown || matches!(progress.progress_state, crate::heal::progress::HealProgressState::Unknown);
+            match stage_current.checked_add(progress.stage_current) {
+                Some(sum) => stage_current = sum,
+                None => counter_overflow = true,
+            }
+            match stage_total.checked_add(progress.stage_total) {
+                Some(sum) => stage_total = sum,
+                None => counter_overflow = true,
+            }
+            for (target, value) in [
+                (&mut snapshot.objects_scanned, progress.objects_scanned),
+                (&mut snapshot.objects_healed, progress.objects_healed),
+                (&mut snapshot.objects_failed, progress.objects_failed),
+                (&mut snapshot.skipped_objects, progress.skipped_objects),
+                (&mut snapshot.skipped_new_versions, progress.skipped_new_versions),
+                (&mut snapshot.skipped_ilm_expired, progress.skipped_ilm_expired),
+                (&mut snapshot.objects_total_count, progress.objects_total_count),
+                (&mut snapshot.objects_total_size, progress.objects_total_size),
+                (&mut snapshot.bytes_processed, progress.bytes_processed),
+            ] {
+                match target.checked_add(value) {
+                    Some(sum) => *target = sum,
+                    None => counter_overflow = true,
+                }
+            }
             snapshot.start_time = match (snapshot.start_time, progress.start_time) {
                 (Some(current), Some(next)) => Some(current.min(next)),
                 (None, next) => next,
@@ -2036,7 +2064,36 @@ impl HealManager {
                 snapshot.current_object = progress.current_object;
             }
         }
-        snapshot.refresh_progress_percentage();
+        snapshot.kind = if has_object_sweep {
+            crate::heal::progress::HealProgressKind::ObjectSweep
+        } else {
+            crate::heal::progress::HealProgressKind::Stage
+        };
+        snapshot.stage_current = stage_current;
+        snapshot.stage_total = stage_total;
+        snapshot.baseline_known = has_object_sweep && all_object_baselines_known;
+        snapshot.progress_state = if counter_overflow {
+            crate::heal::progress::HealProgressState::Unknown
+        } else if has_object_sweep && !all_object_baselines_known {
+            crate::heal::progress::HealProgressState::Indeterminate
+        } else if has_object_sweep {
+            crate::heal::progress::HealProgressState::Running
+        } else if stage_total == 0 {
+            crate::heal::progress::HealProgressState::Indeterminate
+        } else {
+            crate::heal::progress::HealProgressState::Running
+        };
+        if counter_overflow {
+            snapshot.progress_percentage = 0.0;
+        } else if !has_object_sweep {
+            snapshot.progress_percentage = if stage_total == 0 {
+                0.0
+            } else {
+                ((stage_current as f64 / stage_total as f64) * 100.0).min(99.999)
+            };
+        } else {
+            snapshot.refresh_progress_percentage();
+        }
         snapshot.refresh_estimated_completion_time();
         Some(snapshot)
     }

crates/heal/src/heal/progress.rs

@@ -15,15 +15,70 @@
 use serde::{Deserialize, Serialize};
 use std::time::{Duration, SystemTime};
 
+pub(crate) fn increment_counter(counter: &mut u64) -> bool {
+    match counter.checked_add(1) {
+        Some(next) => {
+            *counter = next;
+            true
+        }
+        None => {
+            *counter = u64::MAX;
+            false
+        }
+    }
+}
+
+pub(crate) fn add_bytes(total: &mut u64, amount: u64) -> bool {
+    match total.checked_add(amount) {
+        Some(next) => {
+            *total = next;
+            true
+        }
+        None => {
+            *total = u64::MAX;
+            false
+        }
+    }
+}
+
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub enum HealProgressKind {
+    #[default]
+    Unknown,
+    Stage,
+    ObjectSweep,
+}
+
+/// Whether the object ledger can produce a meaningful percentage.
+///
+/// A zero-valued baseline is not a completed scan: it means that no complete
+/// usage snapshot was available.  Keep this state explicit so callers do not
+/// mistake the legacy `0.0` wire value for a measured zero-percent result.
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub enum HealProgressState {
+    #[default]
+    Unknown,
+    Indeterminate,
+    Running,
+    Completed,
+}
+
 #[derive(Debug, Default, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct HealProgress {
+    #[serde(default)]
+    pub kind: HealProgressKind,
     /// Objects scanned
     pub objects_scanned: u64,
     /// Objects healed
     pub objects_healed: u64,
     /// Objects failed
     pub objects_failed: u64,
+    /// Versions deferred for a later retry pass.
+    #[serde(default)]
+    pub skipped_objects: u64,
     /// Versions skipped because they were written after this heal started
     pub skipped_new_versions: u64,
     /// Versions skipped because lifecycle already selected them for expiry
@@ -44,11 +99,38 @@ pub struct HealProgress {
     pub last_update_time: Option<SystemTime>,
     /// Estimated completion time
     pub estimated_completion_time: Option<SystemTime>,
+    /// Current stage number. Stage updates are intentionally independent from
+    /// the object ledger below.
+    #[serde(default)]
+    pub stage_current: u64,
+    /// Number of stages in the current task.
+    #[serde(default)]
+    pub stage_total: u64,
+    /// Explicitly distinguishes a missing usage baseline from measured 0%.
+    #[serde(default)]
+    pub progress_state: HealProgressState,
+    /// True only after the task's durable completion ledger was committed.
+    #[serde(default)]
+    pub ledger_complete: bool,
+    /// Generation of the usage snapshot used for the baseline, if available.
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    /// Whether the baseline was explicitly observed.  This is separate from
+    /// the counters so a known empty scope (0 objects, 0 bytes) is not
+    /// confused with a legacy snapshot that omitted the baseline fields.
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Internal telemetry fence set when an aggregate counter overflows or
+    /// becomes inconsistent.  It prevents a later refresh from fabricating a
+    /// percentage from the poisoned values.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }
 
 impl HealProgress {
     pub fn new() -> Self {
         Self {
+            kind: HealProgressKind::Unknown,
             start_time: Some(SystemTime::now()),
             last_update_time: Some(SystemTime::now()),
             ..Default::default()
@@ -56,12 +138,87 @@ impl HealProgress {
     }
 
     pub fn update_progress(&mut self, scanned: u64, healed: u64, failed: u64, bytes: u64) {
+        self.update_object_sweep_progress(scanned, healed, failed, bytes);
+    }
+
+    pub fn update_object_sweep_progress(&mut self, scanned: u64, healed: u64, failed: u64, bytes: u64) {
+        self.kind = HealProgressKind::ObjectSweep;
         self.objects_scanned = scanned;
         self.objects_healed = healed;
         self.objects_failed = failed;
         self.bytes_processed = bytes;
         self.last_update_time = Some(SystemTime::now());
 
+        let explicit_skipped = match self.skipped_new_versions.checked_add(self.skipped_ilm_expired) {
+            Some(value) => value,
+            None => {
+                self.mark_unknown();
+                0
+            }
+        };
+        let skipped = healed
+            .checked_add(failed)
+            .and_then(|value| value.checked_add(explicit_skipped))
+            .and_then(|value| scanned.checked_sub(value))
+            .unwrap_or(0);
+        self.update_object_progress(scanned, healed, failed, skipped, bytes);
+    }
+
+    /// Update task stage progress without modifying object counters.
+    pub fn update_stage(&mut self, current: u64, total: u64) {
+        let object_sweep_active = matches!(self.kind, HealProgressKind::ObjectSweep);
+        if !object_sweep_active {
+            self.kind = HealProgressKind::Stage;
+        }
+        self.ledger_complete = false;
+        self.stage_current = current.min(total);
+        self.stage_total = total;
+        if object_sweep_active {
+            self.last_update_time = Some(SystemTime::now());
+            self.refresh_progress_percentage();
+            return;
+        }
+        self.progress_state = if total == 0 {
+            HealProgressState::Indeterminate
+        } else {
+            HealProgressState::Running
+        };
+        self.progress_percentage = if total == 0 {
+            0.0
+        } else {
+            (current as f64 / total as f64 * 100.0).min(100.0)
+        };
+        self.last_update_time = Some(SystemTime::now());
+    }
+
+    /// Update the disjoint object ledger. `scanned` is the number of terminal
+    /// object outcomes and must equal healed + failed + deferred skipped plus
+    /// the two terminal skip classes. Overflow is a corrupt/unknown counter
+    /// state, not a reason to abort a completed heal.
+    pub fn update_object_progress(&mut self, scanned: u64, healed: u64, failed: u64, skipped: u64, bytes: u64) {
+        self.kind = HealProgressKind::ObjectSweep;
+        // `skipped` is the transient/deferred class.  The two explicit skip
+        // counters are terminal classifications too, so include them in the
+        // same ledger without making callers maintain a second aggregate.
+        let outcomes = healed
+            .checked_add(failed)
+            .and_then(|value| value.checked_add(skipped))
+            .and_then(|value| value.checked_add(self.skipped_new_versions))
+            .and_then(|value| value.checked_add(self.skipped_ilm_expired));
+        self.objects_scanned = scanned;
+        self.objects_healed = healed;
+        self.objects_failed = failed;
+        self.skipped_objects = skipped;
+        self.bytes_processed = bytes;
+        self.last_update_time = Some(SystemTime::now());
+        self.ledger_complete = false;
+        if outcomes != Some(scanned) {
+            // Telemetry corruption must not abort a heal.  Preserve the
+            // counters for diagnostics, but do not derive a percentage from a
+            // double-counted or overflowing ledger.
+            self.mark_unknown();
+            return;
+        }
         self.refresh_progress_percentage();
         self.refresh_estimated_completion_time();
     }
@@ -69,50 +226,88 @@ impl HealProgress {
     pub fn set_total_baseline(&mut self, objects_total_count: u64, objects_total_size: u64) {
         self.objects_total_count = objects_total_count;
         self.objects_total_size = objects_total_size;
+        self.baseline_known = true;
         self.last_update_time = Some(SystemTime::now());
         self.refresh_progress_percentage();
         self.refresh_estimated_completion_time();
     }
 
+    pub fn set_total_baseline_with_generation(&mut self, objects_total_count: u64, objects_total_size: u64, generation: u64) {
+        self.baseline_generation = Some(generation);
+        self.set_total_baseline(objects_total_count, objects_total_size);
+    }
+
     pub fn record_skipped_new_version(&mut self) {
-        self.skipped_new_versions = self.skipped_new_versions.saturating_add(1);
+        let Some(next) = self.skipped_new_versions.checked_add(1) else {
+            self.mark_unknown();
+            return;
+        };
+        self.skipped_new_versions = next;
         self.last_update_time = Some(SystemTime::now());
         self.refresh_progress_percentage();
         self.refresh_estimated_completion_time();
     }
 
     pub fn record_skipped_ilm_expired(&mut self) {
-        self.skipped_ilm_expired = self.skipped_ilm_expired.saturating_add(1);
+        let Some(next) = self.skipped_ilm_expired.checked_add(1) else {
+            self.mark_unknown();
+            return;
+        };
+        self.skipped_ilm_expired = next;
         self.last_update_time = Some(SystemTime::now());
         self.refresh_progress_percentage();
         self.refresh_estimated_completion_time();
     }
 
-    fn completed_for_baseline(&self) -> u64 {
+    fn completed_for_baseline(&self) -> Option<u64> {
         self.objects_healed
-            .saturating_add(self.objects_failed)
-            .saturating_add(self.skipped_new_versions)
-            .saturating_add(self.skipped_ilm_expired)
+            .checked_add(self.objects_failed)?
+            .checked_add(self.skipped_objects)?
+            .checked_add(self.skipped_new_versions)?
+            .checked_add(self.skipped_ilm_expired)
     }
 
     pub(crate) fn refresh_progress_percentage(&mut self) {
+        if self.ledger_complete {
+            self.progress_state = HealProgressState::Completed;
+            self.progress_percentage = 100.0;
+            return;
+        }
+        if self.counter_unknown {
+            self.progress_state = HealProgressState::Unknown;
+            self.progress_percentage = 0.0;
+            return;
+        }
+        if !self.baseline_known {
+            self.progress_state = HealProgressState::Indeterminate;
+            self.progress_percentage = 0.0;
+            self.estimated_completion_time = None;
+            return;
+        }
         if self.objects_total_size > 0 {
             self.progress_percentage = ((self.bytes_processed as f64 / self.objects_total_size as f64) * 100.0).min(100.0);
+            self.progress_percentage = self.progress_percentage.min(99.999);
+            self.progress_state = HealProgressState::Running;
             return;
         }
         if self.objects_total_count > 0 {
-            let completed = self.completed_for_baseline();
+            let Some(completed) = self.completed_for_baseline() else {
+                self.progress_state = HealProgressState::Unknown;
+                self.progress_percentage = 0.0;
+                return;
+            };
             self.progress_percentage = ((completed as f64 / self.objects_total_count as f64) * 100.0).min(100.0);
+            self.progress_percentage = self.progress_percentage.min(99.999);
+            self.progress_state = HealProgressState::Running;
             return;
         }
-
-        let total = self
-            .objects_scanned
-            .saturating_add(self.objects_healed)
-            .saturating_add(self.objects_failed);
-        if total > 0 {
-            self.progress_percentage = (self.objects_healed as f64 / total as f64) * 100.0;
+        if self.baseline_known {
+            self.progress_state = HealProgressState::Running;
+            self.progress_percentage = 0.0;
+            return;
         }
+        self.progress_state = HealProgressState::Indeterminate;
+        self.progress_percentage = 0.0;
     }
 
     pub fn set_current_object(&mut self, object: Option<String>) {
@@ -125,7 +320,11 @@ impl HealProgress {
             self.estimated_completion_time = None;
             return;
         };
-        if self.is_completed() || !(0.0..100.0).contains(&self.progress_percentage) || self.bytes_processed == 0 {
+        if self.is_completed()
+            || self.progress_percentage <= 0.0
+            || self.progress_percentage >= 100.0
+            || self.bytes_processed == 0
+        {
             self.estimated_completion_time = None;
             return;
         }
@@ -142,18 +341,39 @@ impl HealProgress {
     }
 
     pub fn is_completed(&self) -> bool {
-        if self.progress_percentage >= 100.0 {
-            return true;
-        }
-        if self.objects_total_count > 0 || self.objects_total_size > 0 {
-            return false;
-        }
+        self.ledger_complete
+    }
 
-        self.objects_scanned > 0 && self.objects_healed.saturating_add(self.objects_failed) >= self.objects_scanned
+    /// Mark telemetry unknown while allowing the underlying heal operation to
+    /// continue.  This is used for corrupt/overflowing counters at the
+    /// observability boundary; it must never turn a successful heal into an
+    /// execution error.
+    pub fn mark_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.progress_state = HealProgressState::Unknown;
+        self.ledger_complete = false;
+        self.progress_percentage = 0.0;
+        self.estimated_completion_time = None;
+        self.last_update_time = Some(SystemTime::now());
+    }
+
+    /// Mark the object ledger terminal only after the enclosing task has
+    /// committed all durable resume state and cleanup fences.
+    pub fn mark_completed(&mut self) {
+        let telemetry_unknown = self.counter_unknown || self.progress_state == HealProgressState::Unknown;
+        self.ledger_complete = true;
+        if !telemetry_unknown {
+            self.progress_state = HealProgressState::Completed;
+        }
+        self.progress_percentage = 100.0;
+        self.last_update_time = Some(SystemTime::now());
+        self.estimated_completion_time = None;
     }
 
     pub fn get_success_rate(&self) -> f64 {
-        let total = self.objects_healed + self.objects_failed;
+        let Some(total) = self.objects_healed.checked_add(self.objects_failed) else {
+            return 0.0;
+        };
         if total > 0 {
             (self.objects_healed as f64 / total as f64) * 100.0
         } else {
@@ -230,6 +450,7 @@ mod tests {
         assert_eq!(progress.objects_scanned, 0);
         assert_eq!(progress.objects_healed, 0);
         assert_eq!(progress.objects_failed, 0);
+        assert_eq!(progress.skipped_objects, 0);
         assert_eq!(progress.skipped_new_versions, 0);
         assert_eq!(progress.skipped_ilm_expired, 0);
         assert_eq!(progress.objects_total_count, 0);
@@ -250,10 +471,8 @@ mod tests {
         assert_eq!(progress.objects_healed, 8);
         assert_eq!(progress.objects_failed, 2);
         assert_eq!(progress.bytes_processed, 1024);
-        // Progress percentage should be calculated based on healed/total
-        // total = scanned + healed + failed = 10 + 8 + 2 = 20
-        // healed/total = 8/20 = 0.4 = 40%
-        assert!((progress.progress_percentage - 40.0).abs() < 0.001);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(progress.progress_percentage, 0.0);
         assert!(progress.last_update_time.is_some());
     }
 
@@ -262,7 +481,8 @@ mod tests {
         let mut progress = HealProgress::new();
         progress.start_time = Some(SystemTime::now() - Duration::from_secs(10));
 
-        progress.update_progress(100, 25, 0, 4096);
+        progress.set_total_baseline(100, 16384);
+        progress.update_progress(25, 25, 0, 4096);
 
         let eta = progress
             .estimated_completion_time
@@ -275,7 +495,7 @@ mod tests {
         let mut progress = HealProgress::new();
         progress.set_total_baseline(10, 8192);
 
-        progress.update_progress(100, 25, 0, 4096);
+        progress.update_progress(25, 25, 0, 4096);
 
         assert!((progress.progress_percentage - 50.0).abs() < 0.001);
     }
@@ -285,7 +505,7 @@ mod tests {
         let mut progress = HealProgress::new();
         progress.set_total_baseline(10, 0);
 
-        progress.update_progress(100, 3, 2, 0);
+        progress.update_progress(5, 3, 2, 0);
 
         assert!((progress.progress_percentage - 50.0).abs() < 0.001);
     }
@@ -295,7 +515,7 @@ mod tests {
         let mut progress = HealProgress::new();
         progress.set_total_baseline(10, 0);
 
-        progress.update_progress(100, 3, 2, 0);
+        progress.update_progress(5, 3, 2, 0);
         progress.record_skipped_new_version();
 
         assert_eq!(progress.skipped_new_versions, 1);
@@ -336,7 +556,8 @@ mod tests {
     fn test_heal_progress_update_progress_all_healed() {
         let mut progress = HealProgress::new();
         // When scanned=0, healed=10, failed=0: total=10, progress = 10/10 = 100%
-        progress.update_progress(0, 10, 0, 2048);
+        progress.update_progress(10, 10, 0, 2048);
+        progress.mark_completed();
 
         // All healed, should be 100%
         assert!((progress.progress_percentage - 100.0).abs() < 0.001);
@@ -394,6 +615,7 @@ mod tests {
         assert_eq!(json["objectsScanned"], 10);
         assert_eq!(json["objectsHealed"], 8);
         assert_eq!(json["objectsFailed"], 2);
+        assert_eq!(json["skippedObjects"], 0);
         assert_eq!(json["skippedNewVersions"], 0);
         assert_eq!(json["skippedIlmExpired"], 0);
         assert_eq!(json["bytesProcessed"], 1024);
@@ -405,6 +627,7 @@ mod tests {
     fn test_heal_progress_is_completed_by_percentage() {
         let mut progress = HealProgress::new();
         progress.update_progress(10, 10, 0, 1024);
+        progress.mark_completed();
 
         assert!(progress.is_completed());
     }
@@ -415,7 +638,7 @@ mod tests {
         progress.objects_scanned = 10;
         progress.objects_healed = 8;
         progress.objects_failed = 2;
-        // healed + failed = 8 + 2 = 10 >= scanned = 10
+        progress.mark_completed();
         assert!(progress.is_completed());
     }
 
@@ -455,6 +678,66 @@ mod tests {
         assert!((progress.get_success_rate() - 100.0).abs() < 0.001);
     }
 
+    #[test]
+    fn single_object_progress_reaches_terminal_100() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert!(!progress.is_completed());
+        progress.mark_completed();
+        assert!(progress.is_completed());
+        assert_eq!(progress.progress_percentage, 100.0);
+    }
+
+    #[test]
+    fn progress_without_baseline_is_indeterminate() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(progress.progress_percentage, 0.0);
+        assert!(progress.estimated_completion_time.is_none());
+    }
+
+    #[test]
+    fn progress_retry_is_exactly_once() {
+        let mut progress = HealProgress::new();
+        progress.set_total_baseline(1, 128);
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert_eq!(progress.objects_scanned, 1);
+        assert_eq!(progress.objects_healed, 1);
+        assert_eq!(progress.bytes_processed, 128);
+    }
+
+    #[test]
+    fn progress_never_triggers_cleanup_before_terminal_ledger_empty() {
+        let mut progress = HealProgress::new();
+        progress.progress_percentage = 100.0;
+        assert!(!progress.is_completed());
+        progress.mark_completed();
+        assert!(progress.is_completed());
+    }
+
+    #[test]
+    fn progress_counter_overflow_is_marked_unknown_without_aborting_completed_heal() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(u64::MAX, u64::MAX, 1, 0, 0);
+        assert_eq!(progress.progress_state, HealProgressState::Unknown);
+        progress.mark_completed();
+        assert!(progress.is_completed());
+        assert_eq!(progress.progress_state, HealProgressState::Unknown);
+    }
+
+    #[test]
+    fn stage_updates_do_not_double_count_object_outcomes() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(2, 1, 0, 1, 256);
+        progress.update_stage(3, 4);
+        assert_eq!(progress.kind, HealProgressKind::ObjectSweep);
+        assert_eq!(progress.objects_scanned, 2);
+        assert_eq!(progress.objects_healed, 1);
+        assert_eq!(progress.skipped_objects, 1);
+    }
+
     #[test]
     fn test_heal_statistics_new() {
         let stats = HealStatistics::new();

crates/heal/src/heal/resume.rs

@@ -31,7 +31,7 @@ mod checkpoint;
 mod replacement;
 mod utils;
 
-pub use checkpoint::{CheckpointManager, ResumeCheckpoint};
+pub use checkpoint::{CheckpointManager, CheckpointObjectOutcome, ResumeCheckpoint};
 pub(crate) use replacement::replacement_target_identities_match;
 use replacement::replacement_targets_match_identities;
 pub use replacement::{
@@ -340,6 +340,12 @@ pub struct ResumeState {
     pub failed_objects: u64,
     /// skipped objects
     pub skipped_objects: u64,
+    /// Terminal versions skipped because they were newer than the heal start.
+    #[serde(default)]
+    pub skipped_new_versions: u64,
+    /// Terminal versions handed to lifecycle expiry.
+    #[serde(default)]
+    pub skipped_ilm_expired: u64,
     /// current bucket
     pub current_bucket: Option<String>,
     /// current object
@@ -354,6 +360,24 @@ pub struct ResumeState {
     pub retry_count: u32,
     /// max retries
     pub max_retries: u32,
+    /// Bytes accounted by the object ledger; additive for old snapshots.
+    #[serde(default)]
+    pub processed_bytes: u64,
+    /// Total bytes from a complete usage snapshot, when available.
+    #[serde(default)]
+    pub total_bytes: u64,
+    /// Generation of the usage snapshot used for the baseline.
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    /// Whether the usage baseline is known. Missing in old snapshots means
+    /// indeterminate rather than a measured zero baseline.
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Persistent telemetry fence for counter/byte overflow or corruption.
+    /// It must survive a restart so a saturated snapshot is never presented as
+    /// a measured percentage on the next resume.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }
 
 impl ResumeState {
@@ -377,6 +401,8 @@ impl ResumeState {
             successful_objects: 0,
             failed_objects: 0,
             skipped_objects: 0,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
             current_bucket: None,
             current_object: None,
             completed_buckets: Vec::new(),
@@ -384,6 +410,11 @@ impl ResumeState {
             error_message: None,
             retry_count: 0,
             max_retries: 3,
+            processed_bytes: 0,
+            total_bytes: 0,
+            baseline_generation: None,
+            baseline_known: false,
+            counter_unknown: false,
         }
     }
 
@@ -412,6 +443,39 @@ impl ResumeState {
         self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
     }
 
+    pub fn update_progress_with_bytes(
+        &mut self,
+        processed: u64,
+        successful: u64,
+        failed: u64,
+        skipped: u64,
+        processed_bytes: u64,
+    ) {
+        self.update_progress(processed, successful, failed, skipped);
+        self.processed_bytes = processed_bytes;
+    }
+
+    pub fn set_skipped_version_counts(&mut self, new_versions: u64, ilm_expired: u64) {
+        self.skipped_new_versions = new_versions;
+        self.skipped_ilm_expired = ilm_expired;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn set_progress_baseline(&mut self, total_objects: u64, total_bytes: u64, generation: Option<u64>) {
+        self.total_objects = total_objects;
+        self.total_bytes = total_bytes;
+        self.baseline_generation = generation;
+        // This method is called only after a complete usage snapshot has been
+        // validated.  A complete but empty snapshot is still a known baseline.
+        self.baseline_known = true;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn mark_counter_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
     pub fn set_current_item(&mut self, bucket: Option<String>, object: Option<String>) {
         self.current_bucket = bucket;
         self.current_object = object;
@@ -437,6 +501,7 @@ impl ResumeState {
         if let Some(pos) = self.pending_buckets.iter().position(|b| b == bucket) {
             self.pending_buckets.remove(pos);
         }
+        self.resume_cursor = None;
         self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
     }
 
@@ -454,6 +519,10 @@ impl ResumeState {
         self.successful_objects = 0;
         self.failed_objects = 0;
         self.skipped_objects = 0;
+        self.skipped_new_versions = 0;
+        self.skipped_ilm_expired = 0;
+        self.processed_bytes = 0;
+        self.counter_unknown = false;
         self.completed = false;
         // A retry re-scans every bucket from the beginning, so the version
         // cursor must be cleared too — otherwise the retry would resume mid-scan.
@@ -476,14 +545,28 @@ impl ResumeState {
     }
 
     pub fn get_progress_percentage(&self) -> f64 {
+        if self.completed {
+            return 100.0;
+        }
+        if self.counter_unknown {
+            return 0.0;
+        }
+        if !self.baseline_known {
+            return 0.0;
+        }
+        if self.total_bytes > 0 {
+            return ((self.processed_bytes as f64 / self.total_bytes as f64) * 100.0).min(99.999);
+        }
         if self.total_objects == 0 {
             return 0.0;
         }
-        (self.processed_objects as f64 / self.total_objects as f64) * 100.0
+        ((self.processed_objects as f64 / self.total_objects as f64) * 100.0).min(99.999)
     }
 
     pub fn get_success_rate(&self) -> f64 {
-        let total = self.successful_objects + self.failed_objects;
+        let Some(total) = self.successful_objects.checked_add(self.failed_objects) else {
+            return 0.0;
+        };
         if total == 0 {
             return 0.0;
         }
@@ -754,6 +837,14 @@ impl ResumeManager {
             state.successful_objects = 0;
             state.failed_objects = 0;
             state.skipped_objects = 0;
+            state.skipped_new_versions = 0;
+            state.skipped_ilm_expired = 0;
+            state.processed_bytes = 0;
+            state.total_objects = 0;
+            state.total_bytes = 0;
+            state.baseline_generation = None;
+            state.baseline_known = false;
+            state.counter_unknown = false;
             state.completed = false;
             state.completed_buckets.clear();
             state.schema_version = CURRENT_RESUME_SCHEMA;
@@ -838,6 +929,41 @@ impl ResumeManager {
         self.save_state_throttled().await
     }
 
+    pub async fn update_progress_with_bytes(
+        &self,
+        processed: u64,
+        successful: u64,
+        failed: u64,
+        skipped: u64,
+        processed_bytes: u64,
+    ) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.update_progress_with_bytes(processed, successful, failed, skipped, processed_bytes);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
+    pub async fn set_progress_baseline(&self, total_objects: u64, total_bytes: u64, generation: Option<u64>) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.set_progress_baseline(total_objects, total_bytes, generation);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
+    pub async fn mark_counter_unknown(&self) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.mark_counter_unknown();
+        drop(state);
+        self.save_state().await
+    }
+
+    pub async fn set_skipped_version_counts(&self, new_versions: u64, ilm_expired: u64) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.set_skipped_version_counts(new_versions, ilm_expired);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
     /// Set current item. Called once per healed object, so persistence is
     /// throttled: the in-memory state always updates, but the snapshot is only
     /// written every `PERSIST_EVERY_MUTATIONS` calls or `PERSIST_INTERVAL`.
@@ -882,7 +1008,7 @@ impl ResumeManager {
         let mut state = self.state.write().await;
         state.complete_bucket(bucket);
         drop(state);
-        self.save_state_throttled().await
+        self.save_state().await
     }
 
     /// mark task completed

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

@@ -34,6 +34,13 @@ const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
 /// to the new `compose_key` identities, so a stale checkpoint is discarded.
 pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
 
+#[derive(Debug, Clone, Copy)]
+pub enum CheckpointObjectOutcome {
+    Processed,
+    Failed,
+    Skipped,
+}
+
 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ResumeCheckpoint {
@@ -57,6 +64,30 @@ pub struct ResumeCheckpoint {
     pub failed_objects: HashSet<String>,
     /// skipped objects
     pub skipped_objects: HashSet<String>,
+    /// Aggregate object ledger counters restored alongside the dedup sets.
+    #[serde(default)]
+    pub successful_objects: u64,
+    #[serde(default)]
+    pub failed_object_count: u64,
+    #[serde(default)]
+    pub skipped_object_count: u64,
+    #[serde(default)]
+    pub skipped_new_versions: u64,
+    #[serde(default)]
+    pub skipped_ilm_expired: u64,
+    #[serde(default)]
+    pub processed_bytes: u64,
+    #[serde(default)]
+    pub total_objects: u64,
+    #[serde(default)]
+    pub total_bytes: u64,
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Persistent telemetry fence for counter/byte overflow or corruption.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }
 
 impl ResumeCheckpoint {
@@ -70,6 +101,17 @@ impl ResumeCheckpoint {
             processed_objects: HashSet::new(),
             failed_objects: HashSet::new(),
             skipped_objects: HashSet::new(),
+            successful_objects: 0,
+            failed_object_count: 0,
+            skipped_object_count: 0,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
+            processed_bytes: 0,
+            total_objects: 0,
+            total_bytes: 0,
+            baseline_generation: None,
+            baseline_known: false,
+            counter_unknown: false,
         }
     }
 
@@ -91,6 +133,34 @@ impl ResumeCheckpoint {
         self.skipped_objects.insert(object);
     }
 
+    pub fn update_progress(&mut self, successful: u64, failed: u64, skipped: u64, bytes: u64) {
+        self.successful_objects = successful;
+        self.failed_object_count = failed;
+        self.skipped_object_count = skipped;
+        self.processed_bytes = bytes;
+    }
+
+    pub fn set_progress_baseline(&mut self, total_objects: u64, total_bytes: u64, generation: Option<u64>) {
+        self.total_objects = total_objects;
+        self.total_bytes = total_bytes;
+        self.baseline_generation = generation;
+        // The caller has already validated that this is a complete snapshot;
+        // preserve the distinction between a known empty scope and an old
+        // checkpoint that omitted all baseline fields.
+        self.baseline_known = true;
+    }
+
+    pub fn mark_counter_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.checkpoint_time = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn set_skipped_version_counts(&mut self, new_versions: u64, ilm_expired: u64) {
+        self.skipped_new_versions = new_versions;
+        self.skipped_ilm_expired = ilm_expired;
+        self.checkpoint_time = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
     /// Advance past a fully-processed page: objects below `object_index` are
     /// skipped by position on resume, so the per-object sets no longer need
     /// their entries and would otherwise grow with the whole bucket.
@@ -107,6 +177,17 @@ impl ResumeCheckpoint {
         self.update_position(0, 0);
         self.processed_objects.clear();
         self.skipped_objects.clear();
+        self.successful_objects = 0;
+        self.failed_object_count = 0;
+        self.skipped_object_count = 0;
+        self.skipped_new_versions = 0;
+        self.skipped_ilm_expired = 0;
+        self.processed_bytes = 0;
+        self.total_objects = 0;
+        self.total_bytes = 0;
+        self.baseline_generation = None;
+        self.baseline_known = false;
+        self.counter_unknown = false;
         self.failed_objects.clear();
     }
 }
@@ -185,6 +266,17 @@ impl CheckpointManager {
             checkpoint.processed_objects.clear();
             checkpoint.failed_objects.clear();
             checkpoint.skipped_objects.clear();
+            checkpoint.successful_objects = 0;
+            checkpoint.failed_object_count = 0;
+            checkpoint.skipped_object_count = 0;
+            checkpoint.skipped_new_versions = 0;
+            checkpoint.skipped_ilm_expired = 0;
+            checkpoint.processed_bytes = 0;
+            checkpoint.total_objects = 0;
+            checkpoint.total_bytes = 0;
+            checkpoint.baseline_generation = None;
+            checkpoint.baseline_known = false;
+            checkpoint.counter_unknown = false;
             checkpoint.current_bucket_index = 0;
             checkpoint.current_object_index = 0;
             checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
@@ -225,7 +317,7 @@ impl CheckpointManager {
         self.save_checkpoint_throttled().await
     }
 
-    /// Advance past a completed page and prune the per-object sets, then persist.
+    /// Persist a completed page position while retaining its identities.
     pub async fn complete_page(&self, bucket_index: usize, object_index: usize) -> Result<()> {
         let mut checkpoint = self.checkpoint.write().await;
         checkpoint.complete_page(bucket_index, object_index);
@@ -233,6 +325,35 @@ impl CheckpointManager {
         self.save_checkpoint_throttled().await
     }
 
+    /// Persist the page position while retaining identities until the resume
+    /// cursor is durable.
+    pub async fn advance_page(&self, bucket_index: usize, object_index: usize) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.update_position(bucket_index, object_index);
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    /// Remove the previous page's dedup identities only after its resume cursor
+    /// has been durably exposed.
+    pub async fn prune_completed_page(&self) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.processed_objects.clear();
+        checkpoint.skipped_objects.clear();
+        checkpoint.failed_objects.clear();
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    /// Advance to the next bucket and clear the final page identities after the
+    /// resume state has durably recorded the completed bucket.
+    pub async fn complete_bucket(&self, next_bucket_index: usize) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.complete_page(next_bucket_index, 0);
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
     /// Reset the checkpoint to the start of the scan for a retry, then persist.
     pub async fn reset_for_retry(&self) -> Result<()> {
         let mut checkpoint = self.checkpoint.write().await;
@@ -267,6 +388,62 @@ impl CheckpointManager {
         self.save_checkpoint_if_due().await
     }
 
+    /// Atomically persist an object's dedup identity with its aggregate result.
+    pub async fn record_object_outcome(
+        &self,
+        object: String,
+        outcome: CheckpointObjectOutcome,
+        successful: u64,
+        failed: u64,
+        skipped: u64,
+        bytes: u64,
+        skipped_new_versions: u64,
+        skipped_ilm_expired: u64,
+        counter_unknown: bool,
+    ) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        match outcome {
+            CheckpointObjectOutcome::Processed => checkpoint.add_processed_object(object),
+            CheckpointObjectOutcome::Failed => checkpoint.add_failed_object(object),
+            CheckpointObjectOutcome::Skipped => checkpoint.add_skipped_object(object),
+        }
+        checkpoint.update_progress(successful, failed, skipped, bytes);
+        checkpoint.set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired);
+        if counter_unknown {
+            checkpoint.mark_counter_unknown();
+        }
+        drop(checkpoint);
+        self.save_checkpoint_if_due().await
+    }
+
+    pub async fn update_progress(&self, successful: u64, failed: u64, skipped: u64, bytes: u64) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.update_progress(successful, failed, skipped, bytes);
+        drop(checkpoint);
+        self.save_checkpoint_if_due().await
+    }
+
+    pub async fn set_progress_baseline(&self, total_objects: u64, total_bytes: u64, generation: Option<u64>) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.set_progress_baseline(total_objects, total_bytes, generation);
+        drop(checkpoint);
+        self.save_checkpoint_throttled().await
+    }
+
+    pub async fn mark_counter_unknown(&self) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.mark_counter_unknown();
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    pub async fn set_skipped_version_counts(&self, new_versions: u64, ilm_expired: u64) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.set_skipped_version_counts(new_versions, ilm_expired);
+        drop(checkpoint);
+        self.save_checkpoint_throttled().await
+    }
+
     async fn save_checkpoint_if_due(&self) -> Result<()> {
         let should_save = self.throttle.lock().map(|mut throttle| throttle.record()).unwrap_or(true);
         if !should_save {

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

@@ -1296,6 +1296,7 @@ async fn test_resume_state_progress() {
     assert_eq!(progress, 0.0); // total_objects is 0
 
     state.total_objects = 100;
+    state.baseline_known = true;
     let progress = state.get_progress_percentage();
     assert_eq!(progress, 10.0);
 }
@@ -1475,6 +1476,40 @@ fn test_checkpoint_object_sets_dedupe_and_prune() {
     assert!(checkpoint.failed_objects.is_empty());
 }
 
+#[tokio::test]
+async fn checkpoint_page_commit_keeps_ledger_until_cursor_is_durable() {
+    let (_temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let checkpoint = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+
+    checkpoint
+        .record_object_outcome(
+            "bucket/object:v1".to_string(),
+            CheckpointObjectOutcome::Processed,
+            1,
+            0,
+            0,
+            128,
+            0,
+            0,
+            false,
+        )
+        .await
+        .unwrap();
+    checkpoint.advance_page(0, 1).await.unwrap();
+
+    let reloaded = CheckpointManager::load_from_disk(disk.clone(), &task_id).await.unwrap();
+    let snapshot = reloaded.get_checkpoint().await;
+    assert_eq!(snapshot.current_object_index, 1);
+    assert_eq!(snapshot.successful_objects, 1);
+    assert_eq!(snapshot.processed_bytes, 128);
+    assert!(snapshot.processed_objects.contains("bucket/object:v1"));
+
+    checkpoint.prune_completed_page().await.unwrap();
+    let reloaded = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    assert!(reloaded.get_checkpoint().await.processed_objects.is_empty());
+}
+
 #[test]
 fn test_checkpoint_loads_legacy_vec_format() {
     // Checkpoints written before the HashSet migration stored the object
@@ -1639,6 +1674,120 @@ async fn current_normal_resume_schema_preserves_progress() {
     temp_dir.close().expect("remove schema test directory");
 }
 
+#[test]
+fn progress_checkpoint_restores_bytes_and_generation() {
+    let mut checkpoint = ResumeCheckpoint::new("progress-checkpoint".to_string());
+    checkpoint.set_progress_baseline(9, 4096, Some(77));
+    checkpoint.update_progress(4, 1, 2, 2048);
+    checkpoint.set_skipped_version_counts(3, 1);
+    checkpoint.mark_counter_unknown();
+
+    let restored: ResumeCheckpoint =
+        serde_json::from_slice(&serde_json::to_vec(&checkpoint).expect("serialize checkpoint")).expect("deserialize checkpoint");
+    assert_eq!(restored.processed_bytes, 2048);
+    assert_eq!(restored.total_objects, 9);
+    assert_eq!(restored.total_bytes, 4096);
+    assert_eq!(restored.baseline_generation, Some(77));
+    assert!(restored.baseline_known);
+    assert_eq!(restored.skipped_new_versions, 3);
+    assert_eq!(restored.skipped_ilm_expired, 1);
+    assert!(restored.counter_unknown);
+}
+
+#[test]
+fn old_progress_schema_migrates_missing_fields_to_unknown() {
+    let state = ResumeState::new(
+        "legacy-progress".to_string(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        Vec::new(),
+    );
+    let mut value = serde_json::to_value(state).expect("serialize legacy-compatible state");
+    let object = value.as_object_mut().expect("state must be an object");
+    for field in [
+        "processed_bytes",
+        "total_bytes",
+        "baseline_generation",
+        "baseline_known",
+        "skipped_new_versions",
+        "skipped_ilm_expired",
+    ] {
+        object.remove(field);
+    }
+    object.insert("total_objects".to_string(), serde_json::json!(10));
+    object.insert("processed_objects".to_string(), serde_json::json!(5));
+    let restored: ResumeState = serde_json::from_value(value).expect("deserialize old progress state");
+    assert_eq!(restored.processed_bytes, 0);
+    assert_eq!(restored.total_bytes, 0);
+    assert_eq!(restored.baseline_generation, None);
+    assert!(!restored.baseline_known, "missing baseline must remain unknown");
+    assert_eq!(restored.get_progress_percentage(), 0.0);
+    assert_eq!(restored.skipped_new_versions, 0);
+    assert_eq!(restored.skipped_ilm_expired, 0);
+}
+
+#[test]
+fn progress_counter_unknown_survives_resume_round_trip() {
+    let mut state = ResumeState::new(
+        "overflow-progress".to_string(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        Vec::new(),
+    );
+    state.mark_counter_unknown();
+
+    let restored: ResumeState =
+        serde_json::from_slice(&serde_json::to_vec(&state).expect("serialize resume state")).expect("deserialize resume state");
+    assert!(restored.counter_unknown);
+}
+
+#[tokio::test]
+async fn checkpoint_progress_survives_a_torn_resume_summary_write() {
+    let (_temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let _resume = ResumeManager::new(
+        disk.clone(),
+        task_id.clone(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        vec!["bucket".to_string()],
+    )
+    .await
+    .expect("resume state should persist");
+    let checkpoint = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("checkpoint should persist");
+
+    // This is the ordering used by the erasure-set loop: the checkpoint is
+    // durable before the summary write.  Stop here to model a crash in the
+    // inter-store window and verify that the recovery authority retains the
+    // telemetry fence and bytes.
+    checkpoint
+        .update_progress(3, 0, 0, 1024)
+        .await
+        .expect("checkpoint progress should persist");
+    checkpoint.mark_counter_unknown().await.expect("unknown fence should persist");
+    checkpoint
+        .update_position(0, 3)
+        .await
+        .expect("checkpoint position should persist");
+
+    let restored_checkpoint = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("checkpoint should reload")
+        .get_checkpoint()
+        .await;
+    let restored_resume = ResumeManager::load_from_disk(disk, &task_id)
+        .await
+        .expect("resume summary should reload")
+        .get_state()
+        .await;
+    assert!(restored_checkpoint.counter_unknown);
+    assert_eq!(restored_checkpoint.processed_bytes, 1024);
+    assert_eq!(restored_checkpoint.current_object_index, 3);
+    assert!(!restored_resume.counter_unknown, "summary is intentionally the torn/older store");
+}
+
 #[tokio::test]
 async fn future_resume_and_checkpoint_schemas_are_rejected() {
     let (temp_dir, disk) = schema_test_disk().await;

crates/heal/src/heal/storage.rs

@@ -19,6 +19,8 @@ use base64::engine::general_purpose::URL_SAFE_NO_PAD;
 use rustfs_common::heal_channel::{HealOpts, HealScanMode};
 use rustfs_madmin::heal_commands::HealResultItem;
 use serde::{Deserialize, Serialize};
+use std::collections::hash_map::DefaultHasher;
+use std::hash::{Hash, Hasher};
 use std::sync::Arc;
 use tracing::{debug, error, warn};
 
@@ -34,6 +36,9 @@ pub use super::{HealObjectInfo, HealObjectOptions, HealPutObjReader};
 pub struct HealBucketUsageBaseline {
     pub objects_count: u64,
     pub bytes: u64,
+    /// Stable identity of the validated usage snapshot and selected scope.
+    /// `None` is retained for test/legacy providers that cannot expose one.
+    pub generation: Option<u64>,
 }
 
 pub struct HealLifecycleExpiryContext {
@@ -785,11 +790,30 @@ impl HealStorageAPI for ECStoreHealStorage {
         let mut baseline = HealBucketUsageBaseline::default();
         for bucket in buckets {
             if let Some(usage) = info.buckets_usage.get(bucket) {
-                baseline.objects_count = baseline.objects_count.saturating_add(usage.objects_count);
-                baseline.bytes = baseline.bytes.saturating_add(usage.size);
+                baseline.objects_count = match baseline.objects_count.checked_add(usage.objects_count) {
+                    Some(total) => total,
+                    // A corrupt/overflowing usage snapshot is not a usable
+                    // denominator.  Leave progress indeterminate instead of
+                    // turning saturation into a plausible percentage.
+                    None => return Ok(None),
+                };
+                baseline.bytes = match baseline.bytes.checked_add(usage.size) {
+                    Some(total) => total,
+                    None => return Ok(None),
+                };
             }
         }
 
+        let identity = info.snapshot_identity();
+        let mut hasher = DefaultHasher::new();
+        identity.last_update.hash(&mut hasher);
+        identity.scanner_cycle.hash(&mut hasher);
+        identity.scanner_epoch.hash(&mut hasher);
+        let mut scope = buckets.to_vec();
+        scope.sort_unstable();
+        scope.hash(&mut hasher);
+        baseline.generation = Some(hasher.finish());
+
         Ok(Some(baseline))
     }
 

crates/heal/src/heal/task.rs

@@ -649,7 +649,7 @@ impl HealTask {
 
         let mut progress = self.progress.write().await;
         progress.set_current_object(Some(format!("skipped: {bucket}/{object}")));
-        progress.update_progress(0, 1, 0, 0);
+        progress.update_stage(1, 1);
         Ok(())
     }
 
@@ -733,7 +733,7 @@ impl HealTask {
             "Heal object skipped for data usage cache after transient error"
         );
         let mut progress = self.progress.write().await;
-        progress.update_progress(3, 3, 0, 0);
+        progress.update_stage(3, 3);
         true
     }
 
@@ -757,7 +757,7 @@ impl HealTask {
         );
         let mut progress = self.progress.write().await;
         progress.set_current_object(Some(format!("skipped: {bucket}/{object}")));
-        progress.update_progress(4, 4, 0, 0);
+        progress.update_stage(4, 4);
         true
     }
 
@@ -831,6 +831,10 @@ impl HealTask {
 
         match &result {
             Ok(_) => {
+                // A stage can reach its final step before the durable resume
+                // ledger and cleanup fences commit. Publish terminal 100 only
+                // after the enclosing operation has returned success.
+                self.progress.write().await.mark_completed();
                 let mut status = self.status.write().await;
                 *status = HealTaskStatus::Completed;
                 demote_to_debug_when!(self.heal_type.is_per_object(), info, target: "rustfs::heal::task", {

crates/heal/src/heal/task/heal_bucket.rs

@@ -13,6 +13,7 @@
 // limitations under the License.
 /// bucket/cluster/prefix heal: the recursive bucket-objects sweep and the erasure-set usage baseline
 use super::*;
+use crate::heal::progress::{add_bytes, increment_counter};
 
 impl HealTask {
     pub(super) async fn heal_bucket(&self, bucket: &str) -> Result<()> {
@@ -32,7 +33,7 @@ impl HealTask {
         {
             let mut progress = self.progress.write().await;
             progress.set_current_object(Some(format!("bucket: {bucket}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
         }
 
         // Step 1: Check if bucket exists
@@ -66,7 +67,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
         }
 
         // Step 2: Perform bucket heal using ecstore
@@ -122,7 +123,7 @@ impl HealTask {
 
                 if !self.options.recursive {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
                 Ok(())
             }
@@ -142,7 +143,7 @@ impl HealTask {
                 );
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
                 Err(Error::TaskExecutionFailed {
                     message: format!("Failed to heal bucket {bucket}: {e}"),
@@ -245,6 +246,7 @@ impl HealTask {
         let mut scanned = 0u64;
         let mut healed = 0u64;
         let mut failed = 0u64;
+        let mut skipped = 0u64;
         let mut retryable_failed = 0u64;
         let mut permanent_failed = 0u64;
         let mut bytes = 0u64;
@@ -286,16 +288,14 @@ impl HealTask {
                 let mut retry = Vec::with_capacity(pending.len());
                 for item in pending {
                     self.check_control_flags().await?;
+                    let mut telemetry_unknown = false;
                     let object = item.name.as_str();
-                    if retry_attempt == 0 {
-                        scanned = scanned.saturating_add(1);
-                    }
                     {
                         let mut progress = self.progress.write().await;
                         progress.set_current_object(Some(format!("{bucket}/{object}")));
-                        progress.update_progress(scanned, healed, failed, bytes);
                     }
 
+                    let mut terminal_outcome = true;
                     let error = match self
                         .await_with_control(
                             self.storage
@@ -304,13 +304,13 @@ impl HealTask {
                         .await
                     {
                         Ok((result, None)) => {
-                            healed = healed.saturating_add(1);
-                            bytes = bytes.saturating_add(u64::try_from(result.object_size).unwrap_or_default());
+                            telemetry_unknown |= !increment_counter(&mut healed);
+                            telemetry_unknown |= !add_bytes(&mut bytes, u64::try_from(result.object_size).unwrap_or(u64::MAX));
                             self.record_result_item(result).await;
                             None
                         }
                         Ok((_, Some(err))) if is_missing_object_dir_heal_result(object, &err) => {
-                            healed = healed.saturating_add(1);
+                            telemetry_unknown |= !increment_counter(&mut healed);
                             debug!(
                                 target: "rustfs::heal::task",
                                 event = EVENT_HEAL_BUCKET_RESULT,
@@ -329,6 +329,7 @@ impl HealTask {
 
                     if let Some(err) = error {
                         if Self::should_skip_data_usage_cache_heal_error(bucket, object, &err) {
+                            telemetry_unknown |= !increment_counter(&mut skipped);
                             warn!(
                                 target: "rustfs::heal::task",
                                 event = EVENT_HEAL_BUCKET_RESULT,
@@ -342,6 +343,7 @@ impl HealTask {
                                 "Heal bucket object repair skipped due to transient metadata error"
                             );
                         } else if err.is_recoverable_heal() && retry_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
+                            terminal_outcome = false;
                             debug!(
                                 target: "rustfs::heal::task",
                                 event = EVENT_HEAL_BUCKET_RESULT,
@@ -357,7 +359,7 @@ impl HealTask {
                             );
                             retry.push(item);
                         } else {
-                            failed = failed.saturating_add(1);
+                            telemetry_unknown |= !increment_counter(&mut failed);
                             if err.is_recoverable_heal() {
                                 retryable_failed = retryable_failed.saturating_add(1);
                             } else {
@@ -383,8 +385,19 @@ impl HealTask {
                         }
                     }
 
+                    if terminal_outcome {
+                        telemetry_unknown |= !increment_counter(&mut scanned);
+                    }
+
+                    if !terminal_outcome {
+                        continue;
+                    }
+
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(scanned, healed, failed, bytes);
+                    progress.update_object_progress(scanned, healed, failed, skipped, bytes);
+                    if telemetry_unknown {
+                        progress.mark_unknown();
+                    }
                 }
                 pending = retry;
                 retry_attempt = retry_attempt.saturating_add(1);
@@ -431,7 +444,7 @@ impl HealTask {
         Ok(())
     }
 
-    pub(super) async fn apply_erasure_set_usage_baseline(&self, buckets: &[String]) -> Result<()> {
+    pub(super) async fn apply_erasure_set_usage_baseline(&self, buckets: &[String], set_disk_id: &str) -> Result<()> {
         let baseline = match self
             .await_with_control(self.storage.erasure_set_usage_baseline(buckets))
             .await
@@ -442,9 +455,26 @@ impl HealTask {
             Err(_) => return Ok(()),
         };
 
-        let HealBucketUsageBaseline { objects_count, bytes } = baseline;
+        let HealBucketUsageBaseline {
+            objects_count,
+            bytes,
+            generation,
+        } = baseline;
+        let generation = generation.map(|snapshot_generation| {
+            use std::hash::{Hash, Hasher};
+            let mut hasher = std::collections::hash_map::DefaultHasher::new();
+            snapshot_generation.hash(&mut hasher);
+            set_disk_id.hash(&mut hasher);
+            self.options.pool_index.hash(&mut hasher);
+            self.options.set_index.hash(&mut hasher);
+            hasher.finish()
+        });
         let mut progress = self.progress.write().await;
-        progress.set_total_baseline(objects_count, bytes);
+        if let Some(generation) = generation {
+            progress.set_total_baseline_with_generation(objects_count, bytes, generation);
+        } else {
+            progress.set_total_baseline(objects_count, bytes);
+        }
         Ok(())
     }
 }

crates/heal/src/heal/task/heal_erasure_set.rs

@@ -32,7 +32,7 @@ impl HealTask {
         {
             let mut progress = self.progress.write().await;
             progress.set_current_object(Some(format!("erasure_set: {} ({} buckets)", set_disk_id, buckets.len())));
-            progress.update_progress(0, 4, 0, 0);
+            progress.update_stage(0, 4);
         }
 
         let is_auto_replacement = matches!(self.source, HealRequestSource::AutoHeal) && !self.heal_endpoints.is_empty();
@@ -158,7 +158,7 @@ impl HealTask {
             None
         };
 
-        self.apply_erasure_set_usage_baseline(&buckets).await?;
+        self.apply_erasure_set_usage_baseline(&buckets, &set_disk_id).await?;
 
         let healing_marker = format!("{set_disk_id}:{}", self.id);
         if let Some((disk, resume_manager, _)) = replacement_resume.as_ref() {
@@ -244,7 +244,7 @@ impl HealTask {
                         );
                         {
                             let mut progress = self.progress.write().await;
-                            progress.update_progress(4, 4, 0, 0);
+                            progress.update_stage(4, 4);
                         }
                         return Err(Error::TaskExecutionFailed {
                             message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
@@ -297,7 +297,7 @@ impl HealTask {
                 );
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(4, 4, 0, 0);
+                    progress.update_stage(4, 4);
                 }
                 return Err(Error::TaskExecutionFailed {
                     message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
@@ -307,7 +307,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(1, 4, 0, 0);
+            progress.update_stage(1, 4);
         }
 
         // The rebuilt disks are formatted now: mark them as healing so
@@ -336,7 +336,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(2, 4, 0, 0);
+            progress.update_stage(2, 4);
         }
 
         // Step 3: Heal bucket structure
@@ -420,7 +420,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(3, 4, 0, 0);
+            progress.update_stage(3, 4);
         }
 
         // Step 4: Execute erasure set heal with resume
@@ -463,9 +463,7 @@ impl HealTask {
         };
 
         {
-            let mut progress = self.progress.write().await;
-            let bytes_processed = progress.bytes_processed;
-            progress.update_progress(4, 4, 0, bytes_processed);
+            self.progress.write().await.update_stage(4, 4);
         }
 
         match result {

crates/heal/src/heal/task/heal_metadata.rs

@@ -32,7 +32,7 @@ impl HealTask {
         {
             let mut progress = self.progress.write().await;
             progress.set_current_object(Some(format!("metadata: {bucket}/{object}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
         }
 
         // Step 1: Check if object exists
@@ -74,7 +74,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
         }
 
         // Step 2: Perform metadata heal using ecstore
@@ -122,7 +122,7 @@ impl HealTask {
                     );
                     {
                         let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                     }
                     return Err(Error::TaskExecutionFailed {
                         message: format!("Failed to heal metadata {bucket}/{object}: {e}"),
@@ -145,7 +145,7 @@ impl HealTask {
 
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
                 self.record_result_item(result).await;
                 Ok(())
@@ -167,7 +167,7 @@ impl HealTask {
                 );
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
                 Err(Error::TaskExecutionFailed {
                     message: format!("Failed to heal metadata {bucket}/{object}: {e}"),
@@ -194,7 +194,7 @@ impl HealTask {
         {
             let mut progress = self.progress.write().await;
             progress.set_current_object(Some(format!("ec_decode: {bucket}/{object}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
         }
 
         // Step 1: Check if object exists
@@ -236,7 +236,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
         }
 
         // Step 2: Perform EC decode heal using ecstore
@@ -284,7 +284,7 @@ impl HealTask {
                     );
                     {
                         let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                     }
                     return Err(Error::TaskExecutionFailed {
                         message: format!("Failed to heal EC decode {bucket}/{object}: {e}"),
@@ -309,7 +309,7 @@ impl HealTask {
 
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                 }
                 self.record_result_item(result).await;
                 Ok(())
@@ -331,7 +331,7 @@ impl HealTask {
                 );
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
                 Err(Error::TaskExecutionFailed {
                     message: format!("Failed to heal EC decode {bucket}/{object}: {e}"),

crates/heal/src/heal/task/heal_object.rs

@@ -36,7 +36,7 @@ impl HealTask {
         {
             let mut progress = self.progress.write().await;
             progress.set_current_object(Some(format!("{bucket}/{object}")));
-            progress.update_progress(0, 4, 0, 0);
+            progress.update_stage(0, 4);
         }
 
         // Step 1: Check if object exists and get metadata
@@ -132,7 +132,7 @@ impl HealTask {
 
         {
             let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
         }
 
         // Step 2: directly call ecstore to perform heal
@@ -187,7 +187,7 @@ impl HealTask {
                         );
                         {
                             let mut progress = self.progress.write().await;
-                            progress.update_progress(3, 3, 0, 0);
+                            progress.update_stage(3, 3);
                         }
                         return Ok(());
                     }
@@ -207,7 +207,7 @@ impl HealTask {
 
                     {
                         let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                     }
 
                     if Self::should_return_typed_heal_error(&e) {
@@ -249,7 +249,7 @@ impl HealTask {
 
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                 }
                 self.record_result_item(result).await;
                 Ok(())
@@ -275,7 +275,7 @@ impl HealTask {
                     );
                     {
                         let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                     }
                     return Ok(());
                 }
@@ -295,7 +295,7 @@ impl HealTask {
 
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                 }
 
                 if Self::should_return_typed_heal_error(&e) {
@@ -414,7 +414,7 @@ impl HealTask {
 
                 {
                     let mut progress = self.progress.write().await;
-                    progress.update_progress(4, 4, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                 }
                 self.record_result_item(result).await;
                 Ok(())

crates/heal/src/heal/task/tests.rs

@@ -2096,6 +2096,7 @@ async fn erasure_set_heal_applies_usage_baseline_to_progress() {
         usage_baseline: Mutex::new(Some(HealBucketUsageBaseline {
             objects_count: 10,
             bytes: 8,
+            generation: Some(1),
         })),
         ..Default::default()
     });
@@ -2119,6 +2120,8 @@ async fn erasure_set_heal_applies_usage_baseline_to_progress() {
     let progress = task.get_progress().await;
     assert_eq!(progress.objects_total_count, 10);
     assert_eq!(progress.objects_total_size, 8);
+    assert!(progress.baseline_generation.is_some());
+    assert!(progress.baseline_known);
     assert_eq!(progress.bytes_processed, 2);
     assert!((progress.progress_percentage - 25.0).abs() < 0.001);
 }

crates/kms/tests/vault_ha_failover_live.rs

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

crates/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/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/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/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/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,

scripts/test/vault_ha_kms_live.sh

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