style: restore rebalance formatting

Signed-off-by: houseme <housemecn@gmail.com>

CLAUDE.md

@@ -30,7 +30,8 @@ make build-docker BUILD_OS=ubuntu22.04
 - Crate membership: `Cargo.toml` `[workspace].members`
 - Architecture, layering, crate map: [ARCHITECTURE.md](ARCHITECTURE.md)
 - Migration guardrails & readiness contracts: [docs/architecture/](docs/architecture/README.md)
-- CI gates: `.github/workflows/ci.yml` (source of truth; never copy its steps into docs)
+- CI workflow steps: `.github/workflows/`; event, timeout, and required-status
+  matrix: [docs/testing/ci-gates.md](docs/testing/ci-gates.md)
 - Test-layer taxonomy, per-layer entry commands, serial/nextest rules, flake
   policy: [docs/testing/README.md](docs/testing/README.md)
 - Tier/ILM transition debugging (xl.meta inspection, versionId tracing):

CONTRIBUTING.md

@@ -70,6 +70,8 @@ make pre-pr
 
 > For the full test-layer taxonomy (unit / ecstore black-box / e2e / s3s-e2e / S3 compatibility / chaos / fuzz / bench), each layer's entry command, the naming conventions the migration gate depends on, and the serial/nextest rules, see [docs/testing/README.md](docs/testing/README.md).
 
+> For the event, timeout, required-status, and local reproduction matrix, see [docs/testing/ci-gates.md](docs/testing/ci-gates.md).
+
 ### 🔒 Automated Pre-commit Hooks
 #### What `make pre-commit` and `make pre-pr` actually run
 

Cargo.lock

@@ -1858,9 +1858,9 @@ dependencies = [
 
 [[package]]
 name = "cc"
-version = "1.4.3"
+version = "1.4.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "509591b7bcd67f4ef775afad7662703b4935daaa6ec0e5605cfb1090b32a2b6d"
+checksum = "0ad534f4357a5264cce5019c989cf66a4f0dc4e0d1b1d15f8aacec0ff7360273"
 dependencies = [
  "find-msvc-tools",
  "jobserver",
@@ -2522,12 +2522,6 @@ dependencies = [
  "subtle",
 ]
 
-[[package]]
-name = "cty"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b365fabc795046672053e29c954733ec3b05e4be654ab130fe8f1f94d7051f35"
-
 [[package]]
 name = "curve25519-dalek"
 version = "4.1.3"
@@ -5988,15 +5982,6 @@ version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
-[[package]]
-name = "libmimalloc-sys"
-version = "0.1.49"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
-dependencies = [
- "cc",
- "cty",
-]
-
 [[package]]
 name = "libredox"
 version = "0.1.20"
@@ -6397,14 +6382,6 @@ dependencies = [
  "synstructure 0.13.2",
 ]
 
-[[package]]
-name = "mimalloc"
-version = "0.1.52"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
-dependencies = [
- "libmimalloc-sys",
-]
-
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -9162,13 +9139,11 @@ dependencies = [
  "insta",
  "jiff",
  "libc",
- "libmimalloc-sys",
  "libsystemd",
  "matchit 0.9.2",
  "md-5 0.11.0",
  "metrics",
  "metrics-util",
- "mimalloc",
  "mime_guess",
  "opentelemetry",
  "opentelemetry_sdk",
@@ -9204,6 +9179,8 @@ dependencies = [
  "rustfs-lock",
  "rustfs-log-analyzer",
  "rustfs-madmin",
+ "rustfs-mimalloc",
+ "rustfs-mimalloc-sys",
  "rustfs-notify",
  "rustfs-object-capacity",
  "rustfs-object-data-cache",
@@ -9875,6 +9852,24 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "rustfs-mimalloc"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a406f4aa07084301d485beec873af6dccc8e3f8762da244743df92038b1db1a6"
+dependencies = [
+ "rustfs-mimalloc-sys",
+]
+
+[[package]]
+name = "rustfs-mimalloc-sys"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3051b819175f58445d4c369a72f0ab88149f3885ba8bea2aff3be01f53fe7cd"
+dependencies = [
+ "cc",
+]
+
 [[package]]
 name = "rustfs-notify"
 version = "1.0.0-rc.3"
@@ -12688,6 +12683,7 @@ dependencies = [
  "js-sys",
  "rand 0.10.2",
  "serde_core",
+ "sha1_smol",
  "wasm-bindgen",
 ]
 

Cargo.toml

@@ -350,8 +350,8 @@ russh-sftp = "2.4.0"
 dav-server = "0.11.0"
 
 # Performance Analysis and Memory Profiling
-mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11" }
-libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11", features = ["extended"] }
+rustfs-mimalloc = { version = "0.5.0" }
+rustfs-mimalloc-sys = { version = "0.5.0" }
 hotpath = { version = "0.23.3", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }

crates/ecstore/src/cluster/rpc/peer_rest_client.rs

@@ -1089,7 +1089,9 @@ impl PeerRestClient {
                     .await?
                     .max_decoding_message_size(BACKGROUND_HEAL_STATUS_MAX_MESSAGE_SIZE);
                 let response = match client
-                    .background_heal_status(Request::new(BackgroundHealStatusRequest::default()))
+                    .background_heal_status(Request::new(BackgroundHealStatusRequest {
+                        protocol_version: rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION,
+                    }))
                     .await
                 {
                     Ok(response) => response.into_inner(),

crates/ecstore/src/core/pools.rs

@@ -2026,7 +2026,7 @@ impl PoolMeta {
         self.load_no_lock(pool).await
     }
 
-    async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
+    pub(crate) async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
     where
         S: EcstoreObjectIO,
     {

crates/ecstore/src/core/sets.rs

@@ -988,14 +988,11 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for Sets {
     }
 }
 
-#[async_trait::async_trait]
-impl crate::storage_api_contracts::heal::HealOperations for Sets {
-    type Error = Error;
-    type HealResultItem = HealResultItem;
-    type HealOptions = HealOpts;
-
-    #[tracing::instrument(skip(self))]
-    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+impl Sets {
+    pub(crate) async fn heal_format_with_fence<F>(&self, dry_run: bool, fence_lost: F) -> Result<(HealResultItem, Option<Error>)>
+    where
+        F: Fn() -> bool + Send + Sync,
+    {
         let (disks, init_errs) = init_storage_disks_with_errors(
             &self.endpoints.endpoints,
             &DiskOption {
@@ -1068,6 +1065,9 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
             // Save new formats `format.json` on unformatted disks.
             for (index, (fm, disk)) in tmp_new_formats.iter_mut().zip(disks.iter()).enumerate() {
                 if fm.is_some() && disk.is_some() {
+                    if fence_lost() {
+                        return Ok((res, Some(StorageError::SlowDown)));
+                    }
                     if let Err(err) = save_format_file(disk, fm).await {
                         if let Some(disk) = disk.as_ref() {
                             let _ = disk.close().await;
@@ -1101,6 +1101,18 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
         }
         Ok((res, None))
     }
+}
+
+#[async_trait::async_trait]
+impl crate::storage_api_contracts::heal::HealOperations for Sets {
+    type Error = Error;
+    type HealResultItem = HealResultItem;
+    type HealOptions = HealOpts;
+
+    #[tracing::instrument(skip(self))]
+    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+        self.heal_format_with_fence(dry_run, || false).await
+    }
     #[tracing::instrument(skip(self))]
     async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem> {
         let mut result = HealResultItem {

crates/ecstore/src/io_support/bitrot.rs

@@ -784,6 +784,24 @@ pub(crate) fn create_deferred_bitrot_reader_with_stripe_handle(
 ///
 /// # Returns
 /// A Result containing the BitrotWriterWrapper or an error
+/// Size hint handed to `DiskAPI::create_file` for a bitrot-wrapped shard.
+///
+/// A known length is grown by one checksum per shard so the on-disk file size
+/// matches what the bitrot writer emits. A negative length is the
+/// unknown-size sentinel (`HashReader::SIZE_PRESERVE_LAYER`, used by SSE and
+/// compression) and must be preserved: `RemoteDisk::create_file` forwards it
+/// in the `put_file_stream` query, and the receiver only treats `size > 0` as
+/// a fixed body length when locating the authenticated trailer. Clamping it
+/// to `0` would claim an empty body and misframe the stream. `0` stays `0`
+/// because a genuinely empty object still means an empty body.
+fn bitrot_create_file_size(length: i64, shard_size: usize, checksum_algo: &HashAlgorithm) -> i64 {
+    if length <= 0 {
+        return length;
+    }
+    let length = length as usize;
+    (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
+}
+
 pub async fn create_bitrot_writer(
     is_inline_buffer: bool,
     disk: Option<&DiskStore>,
@@ -796,12 +814,7 @@ pub async fn create_bitrot_writer(
     let writer = if is_inline_buffer {
         CustomWriter::new_inline_buffer()
     } else if let Some(disk) = disk {
-        let length = if length > 0 {
-            let length = length as usize;
-            (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
-        } else {
-            0
-        };
+        let length = bitrot_create_file_size(length, shard_size, &checksum_algo);
 
         let file = disk.create_file("", volume, path, length).await?;
         #[cfg(feature = "hotpath")]
@@ -820,6 +833,25 @@ mod tests {
     use rustfs_rio::ChunkReader;
     use std::collections::VecDeque;
 
+    #[test]
+    fn bitrot_create_file_size_grows_known_length_by_checksums() {
+        // 10 bytes over 4-byte shards = 3 shards, each followed by a 32-byte hash.
+        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::HighwayHash256), 10 + 3 * 32);
+        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::None), 10);
+    }
+
+    #[test]
+    fn bitrot_create_file_size_keeps_empty_and_unknown_distinct() {
+        assert_eq!(bitrot_create_file_size(0, 4, &HashAlgorithm::HighwayHash256), 0);
+        // SSE/compression streams advertise SIZE_PRESERVE_LAYER (-1); the remote
+        // put_file_stream receiver relies on a non-positive size to parse the auth
+        // trailer from the stream tail, so the sentinel must survive untouched.
+        assert_eq!(
+            bitrot_create_file_size(rustfs_rio::HashReader::SIZE_PRESERVE_LAYER, 4, &HashAlgorithm::HighwayHash256),
+            rustfs_rio::HashReader::SIZE_PRESERVE_LAYER
+        );
+    }
+
     struct TestChunkReader {
         chunks: VecDeque<Bytes>,
     }

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

@@ -2124,26 +2124,13 @@ impl SetDisks {
 
             let put_object_size = known_put_object_storage_size(data.size());
             let shard_file_size_raw = erasure.shard_file_size(put_object_size);
-            let is_inline_buffer =
-                storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);
+            let is_inline_buffer = storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);
 
             let collect_stage_timing = rustfs_io_metrics::put_stage_metrics_enabled() || issue3031_diag_enabled();
             let shard_file_size = shard_file_size_raw;
             let shard_size = erasure.shard_size();
             let write_path = classify_put_write_path(is_inline_buffer, put_object_size, fi.erasure.block_size);
             let direct_inline_commit = matches!(write_path, SmallWritePath::Inline);
-            {
-                use std::io::Write;
-                let msg = format!(
-                    "INLINE_DEBUG: bucket={} obj={} size={} shard_fs={} ds={} bs={} inline={} direct={} path={} iblock={} ver={}\n",
-                    bucket, object, put_object_size, shard_file_size_raw, erasure.data_shards, fi.erasure.block_size,
-                    is_inline_buffer, direct_inline_commit, write_path.metric_label(), storage_class_config.inline_block(), opts.versioned
-                );
-                if let Ok(mut f) = std::fs::OpenOptions::new().create(true).append(true).open("/tmp/rustfs_inline_debug.log") {
-                    let _ = f.write_all(msg.as_bytes());
-                }
-                let _ = std::io::stderr().write_all(msg.as_bytes());
-            }
             rustfs_io_metrics::record_put_object_path(write_path.metric_label());
             let writer_setup_stage_start = collect_stage_timing.then(Instant::now);
             let (mut writers, errors) = if direct_inline_commit {

crates/ecstore/src/store/heal.rs

@@ -13,7 +13,12 @@
 // limitations under the License.
 
 use super::*;
+use crate::core::pools::POOL_META_NAME;
+use crate::services::rebalance::{REBAL_META_NAME, RebalStatus};
+use crate::set_disk::get_lock_acquire_timeout;
 use crate::storage_api_contracts::heal::HealOperations as _;
+use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+use rustfs_lock::NamespaceLockGuard;
 use tracing::trace;
 
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
@@ -30,7 +35,119 @@ fn invalid_heal_pool_index(pool_idx: usize, pool_count: usize) -> Error {
     )
 }
 
+#[derive(Debug, Clone, Copy)]
+enum HealFormatPoolSkip {
+    Completed,
+    Retryable,
+}
+
+fn classify_heal_format_pool(
+    pool_idx: usize,
+    pool_cmd_line: &str,
+    pool_meta: &PoolMeta,
+    rebalance_meta: Option<&RebalanceMeta>,
+) -> Option<HealFormatPoolSkip> {
+    let Some(pool) = pool_meta.pools.get(pool_idx) else {
+        return Some(HealFormatPoolSkip::Retryable);
+    };
+
+    if pool.id != pool_idx || pool_cmd_line.is_empty() || pool.cmd_line.is_empty() || pool.cmd_line != pool_cmd_line {
+        return Some(HealFormatPoolSkip::Retryable);
+    }
+
+    if let Some(decommission) = pool.decommission.as_ref() {
+        if decommission.complete {
+            return Some(HealFormatPoolSkip::Completed);
+        }
+        if decommission.failed || decommission.canceled || decommission.queued || pool_meta.is_suspended(pool_idx) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    if let Some(meta) = rebalance_meta {
+        let Some(pool_stats) = meta.pool_stats.get(pool_idx) else {
+            return Some(HealFormatPoolSkip::Retryable);
+        };
+        if pool_stats.info.stopping || (pool_stats.participating && pool_stats.info.status == RebalStatus::Started) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    None
+}
+
+fn heal_format_pool_skip_error(skip: HealFormatPoolSkip) -> Error {
+    match skip {
+        HealFormatPoolSkip::Completed => StorageError::NoHealRequired,
+        HealFormatPoolSkip::Retryable => StorageError::SlowDown,
+    }
+}
+
+fn heal_format_fence_lost_error() -> Error {
+    StorageError::SlowDown
+}
+
 impl ECStore {
+    async fn acquire_heal_format_fence(
+        &self,
+    ) -> Result<(NamespaceLockGuard, NamespaceLockGuard, PoolMeta, Option<RebalanceMeta>)> {
+        let metadata_pool = self
+            .pools
+            .first()
+            .cloned()
+            .ok_or_else(|| Error::other("heal format requires at least one storage pool"))?;
+
+        // Metadata fence order is part of the decommission/rebalance protocol:
+        // pool.bin must always be acquired before rebalance.bin.
+        let pool_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
+        let pool_guard = pool_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+        let rebalance_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
+        let rebalance_guard = rebalance_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut pool_meta = PoolMeta::default();
+        pool_meta.load_no_lock(metadata_pool.clone()).await?;
+        if pool_meta.pools.len() != self.pools.len()
+            || pool_meta.pools.iter().enumerate().any(|(pool_idx, pool)| {
+                pool.id != pool_idx || pool.cmd_line.is_empty() || pool.cmd_line != self.pools[pool_idx].endpoints.cmd_line
+            })
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut rebalance_meta = RebalanceMeta::new();
+        let rebalance_meta = match rebalance_meta
+            .load_with_opts(
+                metadata_pool,
+                ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+        {
+            Ok(()) => Some(rebalance_meta),
+            Err(Error::ConfigNotFound) => None,
+            Err(err) => return Err(err),
+        };
+
+        if rebalance_meta
+            .as_ref()
+            .is_some_and(|meta| meta.pool_stats.len() != self.pools.len())
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        Ok((pool_guard, rebalance_guard, pool_meta, rebalance_meta))
+    }
+
     fn get_pools_for_heal_object(&self, opts: &HealOpts) -> Result<Vec<Arc<Sets>>> {
         match opts.pool {
             Some(pool_idx) => Ok(vec![
@@ -52,9 +169,26 @@ impl ECStore {
         };
 
         let mut count_no_heal = 0;
+        let mut count_completed = 0;
         let mut first_error = None;
-        for pool in self.pools.iter() {
-            let (mut result, err) = pool.heal_format(dry_run).await?;
+        for (pool_idx, pool) in self.pools.iter().enumerate() {
+            let (pool_guard, rebalance_guard, pool_meta, rebalance_meta) = self.acquire_heal_format_fence().await?;
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
+            if let Some(skip) = classify_heal_format_pool(pool_idx, &pool.endpoints.cmd_line, &pool_meta, rebalance_meta.as_ref())
+            {
+                if matches!(skip, HealFormatPoolSkip::Completed) {
+                    count_completed += 1;
+                } else {
+                    first_error.get_or_insert(heal_format_pool_skip_error(skip));
+                }
+                continue;
+            }
+
+            let fence_lost = || pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost();
+            let (mut result, err) = pool.heal_format_with_fence(dry_run, fence_lost).await?;
             if let Some(err) = err {
                 match err {
                     StorageError::NoHealRequired => {
@@ -69,11 +203,18 @@ impl ECStore {
             r.set_count += result.set_count;
             r.before.drives.append(&mut result.before.drives);
             r.after.drives.append(&mut result.after.drives);
+
+            // A lease can be lost after the final write; fail closed before
+            // reporting the pool as successfully healed.
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
         }
         if let Some(err) = first_error {
             return Ok((r, Some(err)));
         }
-        if count_no_heal == self.pools.len() {
+        if count_no_heal + count_completed == self.pools.len() {
             info!(
                 event = EVENT_HEAL_FORMAT_COMPLETED,
                 component = LOG_COMPONENT_ECSTORE,
@@ -302,6 +443,7 @@ mod tests {
     use crate::disk::{DeleteOptions, DiskOption, format::FormatV3, new_disk};
     use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
     use crate::runtime::instance::InstanceContext;
+    use crate::services::rebalance::{RebalanceInfo, RebalanceStats};
     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};
@@ -353,6 +495,164 @@ mod tests {
         }
     }
 
+    fn pool_meta_with_decommission(info: PoolDecommissionInfo) -> PoolMeta {
+        PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: Some(info),
+            }],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn heal_format_pool_state_barriers_are_classified() {
+        let active = pool_meta_with_decommission(PoolDecommissionInfo {
+            start_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &active, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        for info in [
+            PoolDecommissionInfo {
+                failed: true,
+                ..Default::default()
+            },
+            PoolDecommissionInfo {
+                canceled: true,
+                ..Default::default()
+            },
+        ] {
+            assert!(matches!(
+                classify_heal_format_pool(0, "pool-0", &pool_meta_with_decommission(info), None),
+                Some(HealFormatPoolSkip::Retryable)
+            ));
+        }
+
+        let completed = pool_meta_with_decommission(PoolDecommissionInfo {
+            complete: true,
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &completed, None),
+            Some(HealFormatPoolSkip::Completed)
+        ));
+    }
+
+    #[test]
+    fn heal_format_pool_rebalance_barriers_and_identity_are_fail_closed() {
+        let identity_meta = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        let rebalance = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&rebalance)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopping = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                info: RebalanceInfo {
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity_without_decommission = PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: None,
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity_without_decommission, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "", &identity_meta, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &PoolMeta::default(), None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopped = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Stopped,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopped)).is_none());
+
+        let stopping_after_stop = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping_after_stop)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+    }
+
+    #[test]
+    fn skipped_heal_format_pool_is_never_reported_as_success() {
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Retryable),
+            StorageError::SlowDown
+        ));
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Completed),
+            StorageError::NoHealRequired
+        ));
+    }
+
     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();
@@ -889,6 +1189,18 @@ mod tests {
             bucket_fence_registry: std::sync::Arc::default(),
         };
 
+        let err = store
+            .handle_heal_format(false)
+            .await
+            .expect_err("missing pool metadata must fail closed before format writes");
+        assert!(matches!(err, StorageError::SlowDown));
+
+        let pool_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        pool_meta
+            .save(store.pools.clone())
+            .await
+            .expect("pool metadata should be persisted before format heal");
+
         let (result, err) = store
             .handle_heal_format(false)
             .await
@@ -902,5 +1214,22 @@ mod tests {
             .await
             .expect("the later pool should be healed despite the first pool error");
         assert_eq!(healed.erasure.this, recoverable_format.erasure.sets[0][2]);
+
+        let mut completed_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        for status in &mut completed_meta.pools {
+            status.decommission = Some(PoolDecommissionInfo {
+                complete: true,
+                ..Default::default()
+            });
+        }
+        completed_meta
+            .save(store.pools.clone())
+            .await
+            .expect("completed pool metadata should be persisted");
+        let (_, err) = store
+            .handle_heal_format(false)
+            .await
+            .expect("completed pools should be reported as a no-op");
+        assert!(matches!(err, Some(StorageError::NoHealRequired)));
     }
 }

crates/ecstore/src/store/object.rs

@@ -3194,7 +3194,7 @@ impl ECStore {
 
         // Default return value
         let mut del_objects = vec![DeletedObject::default(); objects.len()];
-        let mut accounting = vec![None; objects.len()];
+        let accounting = vec![None; objects.len()];
 
         let mut del_errs = Vec::with_capacity(objects.len());
         for _ in 0..objects.len() {

crates/ecstore/src/store/rebalance/support.rs

@@ -271,7 +271,7 @@ pub(super) fn resolve_latest_object_info_candidates(
             .filter(|candidate| latest_candidate_mod_time(candidate) == Some(latest_mod_time))
             .collect::<Vec<_>>();
 
-        latest_candidates.sort_by(|left, right| right.idx.cmp(&left.idx));
+        latest_candidates.sort_by_key(|candidate| std::cmp::Reverse(candidate.idx));
 
         let Some(winner) = latest_candidates.first() else {
             return Err(Error::ErasureReadQuorum);

crates/heal/src/heal/erasure_healer.rs

@@ -13,10 +13,10 @@
 // limitations under the License.
 
 use crate::heal::{
-    progress::HealProgress,
+    progress::{HealProgress, add_bytes, increment_counter},
     resume::{
-        CheckpointManager, ReplacementTargetIdentity, ResumeManager, ResumeUtils, compose_key,
-        replacement_target_identities_match,
+        CheckpointManager, CheckpointObjectOutcome, CheckpointObjectOutcomeRecord, ReplacementTargetIdentity, ResumeManager,
+        ResumeUtils, compose_key, replacement_target_identities_match,
     },
     storage::{HealStorageAPI, next_heal_listing_token},
     task::{demote_to_debug_when, is_missing_object_dir_heal_result, take_failure_log_sample},
@@ -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 +881,49 @@ 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 (outcome_record, 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();
+                        }
+                        (
+                            CheckpointObjectOutcomeRecord {
+                                object: key,
+                                outcome: CheckpointObjectOutcome::Processed,
+                                successful: progress.objects_healed,
+                                failed: progress.objects_failed,
+                                skipped: progress.skipped_objects,
+                                bytes: progress.bytes_processed,
+                                skipped_new_versions: progress.skipped_new_versions,
+                                skipped_ilm_expired: progress.skipped_ilm_expired,
+                                counter_unknown: progress.counter_unknown,
+                            },
+                            progress.counter_unknown,
+                        )
+                    };
+                    checkpoint_manager.record_object_outcome(outcome_record).await?;
+                    if counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                     }
                     debug!(
                         target: "rustfs::heal::erasure_healer",
@@ -820,15 +955,41 @@ 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 (outcome_record, 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();
+                        }
+                        (
+                            CheckpointObjectOutcomeRecord {
+                                object: key,
+                                outcome: CheckpointObjectOutcome::Processed,
+                                successful: progress.objects_healed,
+                                failed: progress.objects_failed,
+                                skipped: progress.skipped_objects,
+                                bytes: progress.bytes_processed,
+                                skipped_new_versions: progress.skipped_new_versions,
+                                skipped_ilm_expired: progress.skipped_ilm_expired,
+                                counter_unknown: progress.counter_unknown,
+                            },
+                            progress.counter_unknown,
+                        )
+                    };
+                    checkpoint_manager.record_object_outcome(outcome_record).await?;
+                    if counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                     }
                     debug!(
                         target: "rustfs::heal::erasure_healer",
@@ -954,11 +1115,11 @@ 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 +1132,11 @@ 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 +1149,12 @@ 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 +1167,11 @@ 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 +1184,43 @@ 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 (outcome_record, counter_unknown) = {
                     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();
+                    }
+                    (
+                        CheckpointObjectOutcomeRecord {
+                            object: key,
+                            outcome: checkpoint_outcome,
+                            successful: progress.objects_healed,
+                            failed: progress.objects_failed,
+                            skipped: progress.skipped_objects,
+                            bytes: progress.bytes_processed,
+                            skipped_new_versions: progress.skipped_new_versions,
+                            skipped_ilm_expired: progress.skipped_ilm_expired,
+                            counter_unknown: progress.counter_unknown,
+                        },
+                        progress.counter_unknown,
+                    )
+                };
+                checkpoint_manager.record_object_outcome(outcome_record).await?;
+                if counter_unknown {
+                    resume_manager.mark_counter_unknown().await?;
                 }
 
                 if completed_in_page.is_multiple_of(100) {
@@ -1041,16 +1230,22 @@ 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 +1264,6 @@ 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 +1272,66 @@ 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 +1509,8 @@ 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, CheckpointObjectOutcomeRecord, 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 +1650,7 @@ 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 +1687,9 @@ 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 +1780,7 @@ 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 +1789,7 @@ 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 +2167,49 @@ 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 +2399,175 @@ 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(CheckpointObjectOutcomeRecord {
+                object: compose_key("object", Some("v1")),
+                outcome: CheckpointObjectOutcome::Failed,
+                successful: 0,
+                failed: 1,
+                skipped: 0,
+                bytes: 0,
+                skipped_new_versions: 0,
+                skipped_ilm_expired: 0,
+                counter_unknown: 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

@@ -2011,34 +2011,11 @@ impl HealManager {
             return None;
         }
 
-        let mut snapshot = HealProgress::default();
+        let mut progresses = Vec::with_capacity(active_tasks.len());
         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);
-            snapshot.start_time = match (snapshot.start_time, progress.start_time) {
-                (Some(current), Some(next)) => Some(current.min(next)),
-                (None, next) => next,
-                (current, None) => current,
-            };
-            snapshot.last_update_time = match (snapshot.last_update_time, progress.last_update_time) {
-                (Some(current), Some(next)) => Some(current.max(next)),
-                (None, next) => next,
-                (current, None) => current,
-            };
-            if progress.current_object.is_some() {
-                snapshot.current_object = progress.current_object;
-            }
+            progresses.push(task.get_progress().await);
         }
-        snapshot.refresh_progress_percentage();
-        snapshot.refresh_estimated_completion_time();
-        Some(snapshot)
+        crate::heal::progress::aggregate_heal_progress(progresses)
     }
 }
 

crates/heal/src/heal/progress.rs

@@ -15,15 +15,91 @@
 use serde::{Deserialize, Serialize};
 use std::time::{Duration, SystemTime};
 
-#[derive(Debug, Default, Clone, Serialize, Deserialize)]
+pub(crate) fn stable_generation(parts: &[&[u8]]) -> u64 {
+    let mut hash = 0xcbf29ce484222325u64;
+    for part in parts {
+        for byte in (part.len() as u64).to_be_bytes().into_iter().chain(part.iter().copied()) {
+            hash ^= u64::from(byte);
+            hash = hash.wrapping_mul(0x100000001b3);
+        }
+    }
+    hash
+}
+
+#[cfg(test)]
+mod stable_generation_tests {
+    use super::stable_generation;
+
+    #[test]
+    fn stable_generation_has_a_fixed_vector() {
+        assert_eq!(stable_generation(&[b"rustfs", b"heal", b"42"]), 11_007_672_338_488_385_056);
+    }
+}
+
+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, PartialEq, Serialize, Deserialize)]
+#[serde(default, 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 +120,38 @@ 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 +159,87 @@ 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 +247,88 @@ 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 +341,11 @@ 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 +362,39 @@ 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 {
@@ -162,6 +403,101 @@ impl HealProgress {
     }
 }
 
+pub fn aggregate_heal_progress(progresses: impl IntoIterator<Item = HealProgress>) -> Option<HealProgress> {
+    let mut snapshot = HealProgress::default();
+    let mut found = false;
+    let mut has_object_sweep = false;
+    let mut all_object_baselines_known = true;
+    let mut baseline_generation = None;
+    let mut baseline_generation_consistent = true;
+    let mut all_ledgers_complete = true;
+    let mut counter_overflow = false;
+
+    for progress in progresses {
+        found = true;
+        let object_sweep = matches!(progress.kind, HealProgressKind::ObjectSweep);
+        has_object_sweep |= object_sweep;
+        all_ledgers_complete &= progress.ledger_complete;
+        if object_sweep {
+            all_object_baselines_known &= progress.baseline_known;
+            match baseline_generation {
+                None => baseline_generation = Some(progress.baseline_generation),
+                Some(generation) => baseline_generation_consistent &= generation == progress.baseline_generation,
+            }
+        }
+        counter_overflow |= progress.counter_unknown || matches!(progress.progress_state, HealProgressState::Unknown);
+        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),
+            (&mut snapshot.stage_current, progress.stage_current),
+            (&mut snapshot.stage_total, progress.stage_total),
+        ] {
+            match target.checked_add(value) {
+                Some(sum) => *target = sum,
+                None => {
+                    *target = u64::MAX;
+                    counter_overflow = true;
+                }
+            }
+        }
+        snapshot.start_time = match (snapshot.start_time, progress.start_time) {
+            (Some(current), Some(next)) => Some(current.min(next)),
+            (None, next) => next,
+            (current, None) => current,
+        };
+        snapshot.last_update_time = match (snapshot.last_update_time, progress.last_update_time) {
+            (Some(current), Some(next)) => Some(current.max(next)),
+            (None, next) => next,
+            (current, None) => current,
+        };
+        if progress.current_object.is_some() {
+            snapshot.current_object = progress.current_object;
+        }
+    }
+
+    if !found {
+        return None;
+    }
+
+    snapshot.kind = if has_object_sweep {
+        HealProgressKind::ObjectSweep
+    } else {
+        HealProgressKind::Stage
+    };
+    snapshot.baseline_known = has_object_sweep && all_object_baselines_known && baseline_generation_consistent;
+    snapshot.baseline_generation = if snapshot.baseline_known && baseline_generation_consistent {
+        baseline_generation.flatten()
+    } else {
+        None
+    };
+    snapshot.ledger_complete = all_ledgers_complete;
+    snapshot.counter_unknown = counter_overflow;
+    if counter_overflow {
+        snapshot.progress_state = HealProgressState::Unknown;
+        snapshot.progress_percentage = if snapshot.ledger_complete { 100.0 } else { 0.0 };
+    } else if snapshot.ledger_complete {
+        snapshot.progress_state = HealProgressState::Completed;
+        snapshot.progress_percentage = 100.0;
+    } else if has_object_sweep {
+        snapshot.refresh_progress_percentage();
+    } else if snapshot.stage_total == 0 {
+        snapshot.progress_state = HealProgressState::Indeterminate;
+        snapshot.progress_percentage = 0.0;
+    } else {
+        snapshot.progress_state = HealProgressState::Running;
+        snapshot.progress_percentage = ((snapshot.stage_current as f64 / snapshot.stage_total as f64) * 100.0).min(99.999);
+    }
+    snapshot.refresh_estimated_completion_time();
+    Some(snapshot)
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct HealStatistics {
     /// Total heal tasks
@@ -230,6 +566,7 @@ 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 +587,8 @@ 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 +597,8 @@ 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 +611,7 @@ 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 +621,7 @@ 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 +631,7 @@ 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 +672,8 @@ 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 +731,7 @@ 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 +743,7 @@ 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 +754,7 @@ 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 +794,108 @@ 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);
+
+        let aggregate = aggregate_heal_progress([progress]).expect("progress should aggregate");
+        assert!(aggregate.ledger_complete);
+        assert!(aggregate.counter_unknown);
+        assert_eq!(aggregate.progress_state, HealProgressState::Unknown);
+        assert_eq!(aggregate.progress_percentage, 100.0);
+    }
+
+    #[test]
+    fn aggregate_rejects_mixed_baseline_generations() {
+        let progress = |generation| HealProgress {
+            kind: HealProgressKind::ObjectSweep,
+            objects_scanned: 5,
+            objects_total_count: 10,
+            progress_state: HealProgressState::Running,
+            baseline_generation: Some(generation),
+            baseline_known: true,
+            ..Default::default()
+        };
+
+        let aggregate = aggregate_heal_progress([progress(1), progress(2)]).expect("progress should aggregate");
+        assert!(!aggregate.baseline_known);
+        assert_eq!(aggregate.baseline_generation, None);
+        assert_eq!(aggregate.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(aggregate.progress_percentage, 0.0);
+    }
+
+    #[test]
+    fn aggregate_accepts_multiple_sets_from_one_snapshot_generation() {
+        let progress = |objects_scanned| HealProgress {
+            kind: HealProgressKind::ObjectSweep,
+            objects_scanned,
+            objects_total_count: 10,
+            progress_state: HealProgressState::Running,
+            baseline_generation: Some(7),
+            baseline_known: true,
+            ..Default::default()
+        };
+
+        let aggregate = aggregate_heal_progress([progress(5), progress(3)]).expect("progress should aggregate");
+        assert!(aggregate.baseline_known);
+        assert_eq!(aggregate.baseline_generation, Some(7));
+    }
+
+    #[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, CheckpointObjectOutcomeRecord, 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

@@ -29,10 +29,30 @@ use super::{
 
 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
 
-/// Current on-disk schema version for `ResumeCheckpoint`. Same rationale as
-/// `CURRENT_RESUME_SCHEMA`: pre-per-version dedup identities are not comparable
-/// to the new `compose_key` identities, so a stale checkpoint is discarded.
-pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
+/// Current on-disk schema version for `ResumeCheckpoint`. Schema 5 could
+/// persist dedup identities without the aggregate counters needed to restore
+/// them safely, so stale checkpoints are discarded and replayed.
+pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 6;
+
+#[derive(Debug, Clone, Copy)]
+pub enum CheckpointObjectOutcome {
+    Processed,
+    Failed,
+    Skipped,
+}
+
+#[derive(Debug)]
+pub struct CheckpointObjectOutcomeRecord {
+    pub object: String,
+    pub outcome: CheckpointObjectOutcome,
+    pub successful: u64,
+    pub failed: u64,
+    pub skipped: u64,
+    pub bytes: u64,
+    pub skipped_new_versions: u64,
+    pub skipped_ilm_expired: u64,
+    pub counter_unknown: bool,
+}
 
 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -57,6 +77,30 @@ 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 +114,17 @@ 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 +146,34 @@ 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 +190,17 @@ 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();
     }
 }
@@ -158,10 +252,9 @@ impl CheckpointManager {
             });
         }
 
-        // A checkpoint from an older schema stored latest-only dedup identities
-        // that are not comparable to the new per-version `compose_key`
-        // identities. Discard the stale sets and position, then stamp the
-        // current schema so the scan restarts cleanly.
+        // Older checkpoints can contain identities that are not comparable to
+        // the current keys or lack their corresponding aggregate counters.
+        // Discard the stale sets and position so the scan restarts cleanly.
         if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
             return Err(Error::TaskExecutionFailed {
                 message: format!(
@@ -185,6 +278,17 @@ 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 +329,7 @@ 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 +337,35 @@ 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 +400,62 @@ 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, record: CheckpointObjectOutcomeRecord) -> Result<()> {
+        let CheckpointObjectOutcomeRecord {
+            object,
+            outcome,
+            successful,
+            failed,
+            skipped,
+            bytes,
+            skipped_new_versions,
+            skipped_ilm_expired,
+            counter_unknown,
+        } = record;
+        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(CheckpointObjectOutcomeRecord {
+            object: "bucket/object:v1".to_string(),
+            outcome: CheckpointObjectOutcome::Processed,
+            successful: 1,
+            failed: 0,
+            skipped: 0,
+            bytes: 128,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
+            counter_unknown: 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
@@ -1568,14 +1603,14 @@ async fn test_resumestate_schema_v0_discarded_on_load() {
 }
 
 #[tokio::test]
-async fn test_checkpoint_schema_v4_discarded_on_load() {
+async fn test_checkpoint_schema_v5_discarded_on_load() {
     let (temp_dir, disk) = schema_test_disk().await;
 
-    // The previous checkpoint schema is unsafe once its paired resume
-    // state is discarded: retaining either position would skip work.
+    // Schema v5 can persist failed identities without the aggregate counters
+    // that make those identities safe to deduplicate after an upgrade.
     let task_id = "00000000-0000-4000-8000-000000000002";
     let legacy = r#"{
-            "schema_version": 4,
+            "schema_version": 5,
             "task_id": "00000000-0000-4000-8000-000000000002",
             "checkpoint_time": 1700000000,
             "current_bucket_index": 2,
@@ -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

@@ -22,6 +22,7 @@ use serde::{Deserialize, Serialize};
 use std::sync::Arc;
 use tracing::{debug, error, warn};
 
+use super::progress::stable_generation;
 use super::storage_api::owner::{EcstoreHealLifecycleExpiryContext, ecstore_load_admin_data_usage_from_backend_cached};
 use super::storage_api::storage::{
     BucketInfo, BucketOperations, DiskSetSelector, HealOperations as _, ListOperations as _, ObjectIO as _,
@@ -34,6 +35,9 @@ 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 +789,52 @@ 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 canonical = Vec::new();
+        match identity.last_update {
+            Some(last_update) => {
+                canonical.push(1);
+                canonical.extend_from_slice(
+                    &last_update
+                        .duration_since(std::time::UNIX_EPOCH)
+                        .unwrap_or_default()
+                        .as_nanos()
+                        .to_be_bytes(),
+                );
+            }
+            None => canonical.push(0),
+        }
+        for value in [identity.scanner_cycle, identity.scanner_epoch] {
+            match value {
+                Some(value) => {
+                    canonical.push(1);
+                    canonical.extend_from_slice(&value.to_be_bytes());
+                }
+                None => canonical.push(0),
+            }
+        }
+        let mut scope = buckets.to_vec();
+        scope.sort_unstable();
+        for bucket in scope {
+            canonical.extend_from_slice(&(bucket.len() as u64).to_be_bytes());
+            canonical.extend_from_slice(bucket.as_bytes());
+        }
+        baseline.generation = Some(stable_generation(&[&canonical]));
+
         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, stable_generation};
 
 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 +444,10 @@ 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<()> {
+        if matches!(self.options.scan_mode, HealScanMode::Deep) || matches!(self.source, HealRequestSource::AutoHeal) {
+            return Ok(());
+        }
         let baseline = match self
             .await_with_control(self.storage.erasure_set_usage_baseline(buckets))
             .await
@@ -442,9 +458,18 @@ impl HealTask {
             Err(_) => return Ok(()),
         };
 
-        let HealBucketUsageBaseline { objects_count, bytes } = baseline;
+        let HealBucketUsageBaseline {
+            objects_count,
+            bytes,
+            generation,
+        } = baseline;
+        let generation = generation.map(|snapshot_generation| stable_generation(&[&snapshot_generation.to_be_bytes()]));
         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() {
@@ -231,6 +231,10 @@ impl HealTask {
                             "Heal erasure set format repair skipped because no format heal was required"
                         );
                     } else {
+                        let error = e;
+                        if error.is_recoverable_heal() {
+                            return Err(error);
+                        }
                         error!(
                             target: "rustfs::heal::task",
                             event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -239,15 +243,15 @@ impl HealTask {
                             task_id = %self.id,
                             set_disk_id,
                             result = "format_failed",
-                            error = %e,
+                            error = %error,
                             "Heal erasure set failed"
                         );
                         {
                             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}"),
+                            message: format!("Failed to heal disk format for {set_disk_id}: {error}"),
                         });
                     }
                 } else {
@@ -284,6 +288,9 @@ impl HealTask {
             Err(Error::TaskCancelled) => return Err(Error::TaskCancelled),
             Err(Error::TaskTimeout) => return Err(Error::TaskTimeout),
             Err(e) => {
+                if e.is_recoverable_heal() {
+                    return Err(e);
+                }
                 error!(
                     target: "rustfs::heal::task",
                     event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -297,7 +304,7 @@ 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 +314,7 @@ 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 +343,7 @@ 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 +427,7 @@ 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 +470,7 @@ 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

@@ -22,6 +22,7 @@ use std::sync::Mutex;
 use tempfile::TempDir;
 
 use super::super::storage_api::status::BucketInfo;
+use crate::heal::progress::HealProgressState;
 
 #[tokio::test]
 async fn retry_request_carries_remaining_timeout_budget() {
@@ -547,6 +548,7 @@ struct MockStorage {
     heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
     heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
     format_no_heal_required: Mutex<bool>,
+    format_error: Mutex<Option<Error>>,
     global_format_calls: Mutex<u32>,
     replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
     replacement_targets_ready: Mutex<bool>,
@@ -867,6 +869,9 @@ impl HealStorageAPI for MockStorage {
 
     async fn heal_format(&self, _dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
         *self.global_format_calls.lock().unwrap() += 1;
+        if let Some(error) = self.format_error.lock().unwrap().take() {
+            return Err(error);
+        }
         let no_heal_required = *self.format_no_heal_required.lock().unwrap();
         if no_heal_required {
             Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
@@ -2052,6 +2057,30 @@ async fn test_erasure_set_heal_continues_after_format_no_heal_required() {
     );
 }
 
+#[tokio::test]
+async fn erasure_set_format_slowdown_is_propagated() {
+    let storage = Arc::new(MockStorage {
+        format_error: Mutex::new(Some(Error::Storage(EcstoreError::SlowDown))),
+        ..Default::default()
+    });
+    let request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions::default(),
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("format SlowDown must remain recoverable for the task manager");
+
+    assert!(matches!(error, Error::Storage(EcstoreError::SlowDown)));
+}
+
 #[tokio::test]
 async fn erasure_set_bucket_prepass_failure_stops_before_object_heal() {
     let temp = TempDir::new().expect("temporary directory should be created");
@@ -2096,6 +2125,7 @@ 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()
     });
@@ -2119,10 +2149,55 @@ 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);
 }
 
+#[tokio::test]
+async fn erasure_set_disk_walk_keeps_cluster_usage_baseline_indeterminate() {
+    for (scan_mode, source) in [
+        (HealScanMode::Deep, HealRequestSource::Admin),
+        (HealScanMode::Normal, HealRequestSource::AutoHeal),
+    ] {
+        let temp = TempDir::new().expect("temporary directory should be created");
+        let disk = make_resume_disk(&temp).await;
+        let storage = Arc::new(MockStorage {
+            resume_disk: Mutex::new(Some(disk)),
+            usage_baseline: Mutex::new(Some(HealBucketUsageBaseline {
+                objects_count: 10,
+                bytes: 8,
+                generation: Some(1),
+            })),
+            ..Default::default()
+        });
+        let mut request = HealRequest::new(
+            HealType::ErasureSet {
+                buckets: vec!["bucket-a".to_string()],
+                set_disk_id: "pool_0_set_0".to_string(),
+            },
+            HealOptions {
+                scan_mode,
+                timeout: None,
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        );
+        request.source = source;
+        let task = HealTask::from_request(request, storage);
+
+        task.heal_erasure_set(vec!["bucket-a".to_string()], "pool_0_set_0".to_string())
+            .await
+            .expect("erasure set heal should complete");
+
+        let progress = task.get_progress().await;
+        assert!(!progress.baseline_known);
+        assert_eq!(progress.baseline_generation, None);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+    }
+}
+
 #[tokio::test]
 async fn erasure_set_heal_ignores_usage_baseline_errors() {
     let temp = TempDir::new().expect("temporary directory should be created");

crates/heal/src/lib.rs

@@ -19,7 +19,7 @@ pub use error::{Error, Result};
 pub use heal::{
     HealManager, HealOperationsSnapshot, HealOptions, HealPriority, HealPriorityCounts, HealRequest, HealSourceCounts, HealType,
     channel::HealChannelProcessor,
-    progress::HealProgress,
+    progress::{HealProgress, aggregate_heal_progress},
     resume::{ReplacementRecoveryRecord, ReplacementRecoveryState, ResumeUtils},
 };
 use rustfs_concurrency::WorkloadAdmissionSnapshotProvider;

crates/protos/src/generated/proto_gen/node_service.rs

@@ -1215,7 +1215,10 @@ pub struct ScannerActivityResponse {
     pub dirty_usage_pending: bool,
 }
 #[derive(Clone, Copy, PartialEq, Eq, Hash, ::prost::Message)]
-pub struct BackgroundHealStatusRequest {}
+pub struct BackgroundHealStatusRequest {
+    #[prost(uint32, tag = "1")]
+    pub protocol_version: u32,
+}
 #[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
 pub struct BackgroundHealStatusResponse {
     #[prost(bool, tag = "1")]

crates/protos/src/lib.rs

@@ -170,6 +170,7 @@ pub fn internode_rpc_max_message_size() -> usize {
 pub const HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE: usize = heal_control::RESULT_MAX_SIZE + 1024;
 pub const HEAL_CONTROL_PROTOCOL_VERSION: u32 = 3;
 pub const DYNAMIC_CONFIG_PROTOCOL_VERSION: u32 = 1;
+pub const BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION: u32 = 2;
 pub const HEAL_CONTROL_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-heal-control-capability-v3\0";
 pub const REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-tier-remote-version-state-capability-v1\0";
 pub const CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-cross-pool-fence-capability-v1\0";
@@ -2346,10 +2347,28 @@ pub async fn evict_failed_connection_with_log_level(addr: &str, log_level: Conne
 #[cfg(test)]
 mod tests {
     use super::*;
+    use prost::Message as _;
     use std::sync::Mutex;
 
     static INTERNODE_RPC_MSGPACK_ONLY_ENV_LOCK: Mutex<()> = Mutex::new(());
 
+    #[derive(Clone, PartialEq, prost::Message)]
+    struct BackgroundHealStatusRequestV1 {}
+
+    #[test]
+    fn background_heal_status_request_remains_rolling_upgrade_compatible() {
+        let current = proto_gen::node_service::BackgroundHealStatusRequest {
+            protocol_version: BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION,
+        };
+        let encoded = current.encode_to_vec();
+        BackgroundHealStatusRequestV1::decode(encoded.as_slice()).expect("v1 server should ignore the version field");
+
+        let encoded = BackgroundHealStatusRequestV1 {}.encode_to_vec();
+        let decoded = proto_gen::node_service::BackgroundHealStatusRequest::decode(encoded.as_slice())
+            .expect("v2 server should accept a v1 request");
+        assert_eq!(decoded.protocol_version, 0);
+    }
+
     #[derive(Clone, Copy, Debug, Eq, Ord, PartialEq, PartialOrd)]
     struct CompatPayloadField {
         message: &'static str,

crates/protos/src/node.proto

@@ -846,7 +846,9 @@ message ScannerActivityResponse {
   bool dirty_usage_pending = 9;
 }
 
-message BackgroundHealStatusRequest {}
+message BackgroundHealStatusRequest {
+  uint32 protocol_version = 1;
+}
 
 message BackgroundHealStatusResponse {
   bool success = 1;

crates/test-utils/src/lib.rs

@@ -245,6 +245,18 @@ impl TestECStoreEnvBuilder {
             .await
             .expect("build test ECStore");
 
+        // The production bootstrap only persists pool.bin from the elected
+        // first cluster node.  Test stores intentionally have no cluster
+        // election, but heal-format still requires that durable fence before
+        // it can write any disk format.  Materialize the validated topology
+        // here so the shared fixture models a ready single-node store.
+        let mut pool_meta = ecstore.pool_meta.read().await.clone();
+        pool_meta.dont_save = false;
+        pool_meta
+            .save(ecstore.pools.clone())
+            .await
+            .expect("persist test pool metadata");
+
         if self.init_bucket_metadata {
             let buckets_list = ecstore
                 .list_bucket(&BucketOptions {

deny.toml

@@ -43,9 +43,6 @@ allow-git = [
     # RustFS fork carrying presigned expiry and constant-time authentication fixes.
     # owner: rustfs-maintainers  review: 2026-10
     "https://github.com/rustfs/s3s.git",
-    # MiMalloc fork pinned for hotpath allocation counting support.
-    # owner: houseme  review: 2026-10
-    "https://github.com/xonatius/mimalloc_rust.git",
 ]
 
 [bans]

docs/testing/ci-gates.md

@@ -0,0 +1,149 @@
+# CI gate matrix
+
+This file is the source of truth for which validation runs on each event, its
+configured wall-clock budget, and whether it can block a merge. Test taxonomy,
+naming, and nextest serialization rules remain in [README.md](README.md); e2e
+membership and counts remain in
+[e2e-suite-inventory.md](e2e-suite-inventory.md).
+
+The distinction between **required** and **report-only** is load-bearing:
+a failing job blocks a merge only when its exact check name is present in the
+live `main` ruleset. A workflow name, a `merge_group` trigger, or a red PR check
+does not make a job required by itself.
+
+## Required merge checks
+
+The live `main` ruleset (`6436880`) currently requires exactly these contexts:
+
+| Required context | Producer | Validation |
+|---|---|---|
+| `CLA Check` | `.github/workflows/cla.yml` | Contributor agreement |
+| `Quick Checks` | `.github/workflows/ci.yml` | Formatting and repository guard scripts |
+| `Test and Lint` | `.github/workflows/ci.yml` | Clippy, workspace nextest excluding `e2e_test`, doctests, and migration proofs |
+
+For pull requests limited to the paths excluded by the main CI workflow,
+`.github/workflows/ci-docs-only.yml` reports `Quick Checks` and
+`Test and Lint` under the same names. It runs the real quick checks and the
+planning-document guard; it does not claim that Rust compilation or runtime
+tests ran. Despite the workflow name, these paths also include selected deploy,
+workflow, and lock files.
+
+Verify the live rule rather than trusting this snapshot before changing merge
+policy:
+
+```bash
+gh api repos/rustfs/rustfs/rulesets/6436880 \
+  --jq '.rules[] | select(.type == "required_status_checks") | .parameters'
+```
+
+The ruleset currently has `strict_required_status_checks_policy=false`.
+`Continuous Integration` accepts `merge_group` events and runs `e2e-full` for
+them, but `End-to-End Tests (full merge gate)` is not currently a required
+context. Therefore the repository is prepared to test a merge-queue SHA, but
+the workflow alone does not prove that every merge passed that lane.
+
+## Pull request and merge matrix
+
+Budgets below are job `timeout-minutes`, not typical runtimes. “Report-only”
+means the result is visible and actionable but is not in the live required
+context list.
+
+| Event | Validation | Budget | Merge status | Reproduction |
+|---|---|---:|---|---|
+| PR, non-doc change | `Quick Checks` | 10 min | Required | `make pre-commit` (broader local umbrella) |
+| PR, non-doc change | `Test and Lint` | 90 min | Required | `cargo nextest run --profile ci --all --exclude e2e_test` |
+| PR, non-doc change | `Typos` | 10 min | Report-only | `typos` |
+| PR, non-doc change | `ILM Integration (serial)` | 90 min | Report-only | Use the exact command in `.github/workflows/ci.yml` |
+| PR, non-doc change | rio-v2 / swift / sftp test-and-lint variants | 90 min each | Report-only | `cargo nextest run` with the workflow's feature set |
+| PR, non-doc change | `Build RustFS Debug Binary` | 30 min | Report-only; prerequisite for black-box lanes | `cargo build -p rustfs --bins` |
+| PR, non-doc change | `io_uring Integration (real)` | 30 min | Report-only | `cargo test -p rustfs-ecstore --lib uring_ -- --test-threads=1 --nocapture` |
+| PR, non-doc change | `End-to-End Tests` (`e2e-smoke` plus `s3s-e2e`) | 30 min | Report-only | `cargo nextest run --profile e2e-smoke -p e2e_test`; then `./scripts/e2e-run.sh ./target/debug/rustfs <data-dir>` |
+| PR, non-doc change | `S3 Implemented Tests` | 60 min | Report-only | Build `rustfs`, then run `scripts/s3-tests/run.sh` with `DEPLOY_MODE=binary`, `TEST_MODE=single`, and `MAXFAIL=0` |
+| PR, non-doc change | `S3 Lifecycle Behavior Tests` | 30 min | Report-only | Use the accelerated scanner environment in `.github/workflows/ci.yml` with `scripts/s3-tests/run.sh` |
+| PR touching dependency or workflow inputs | Cargo Deny / Workflow Pin Report / Dependency Review | 20 / 5 / 30 min | Report-only | `cargo deny check`; `scripts/security/check_workflow_pins.sh` |
+| PR touching architecture rules or architecture docs | `Architecture Migration Rules` | 10 min | Report-only | `scripts/check_architecture_migration_rules.sh` |
+| PR touching Nix or workspace manifests | `Nix Build & Check` | 60 min | Report-only | `nix flake check` |
+| PR limited to main-CI-excluded paths | companion `Quick Checks` and `Test and Lint` | 10 min each | Required | `git diff --check`; `make doc-paths-check` when documentation paths changed |
+| `merge_group` | Standard CI plus `e2e-full` | 55 min for `e2e-full` | Standard required contexts only; `e2e-full` report-only | `cargo nextest run --profile e2e-full -p e2e_test` |
+| Push to `main` | Standard CI plus `e2e-full` | 55 min for `e2e-full` | Post-merge detection | Same as `merge_group` |
+| PR touching fuzz inputs or harness paths | Build plus five 60-second fuzz smoke targets | 60 min build; 30 min per target | Report-only | `MAX_TOTAL_TIME=60 ./scripts/fuzz/run.sh` |
+| PR touching selected ecstore disk/format paths | `Rename Safety` on Windows | 60 min | Report-only | Run the four `cargo test -p rustfs-ecstore --lib <filter>` commands in `windows-filesystem.yml` on Windows |
+
+The authoritative e2e filters live in `.config/nextest.toml`; extend a profile
+instead of adding a second ad-hoc selector. Before a profile runs,
+`scripts/check_test_wiring.py` compares its exact membership to the committed
+digest so a silent test drop fails closed.
+
+## Scheduled and manual validation
+
+Scheduled lanes are independent fault domains. They do not block a pull
+request, but their workflow-local gate can fail the run and scheduled failures
+are routed to the shared failure-issue action. The scheduled-validation
+watchdog and freshness workflow separately detect incomplete runs and missing
+schedules.
+
+| Cadence (UTC unless noted) | Workflow / validation | Budget | Verdict and artifacts | Reproduction |
+|---|---|---:|---|---|
+| Daily 02:17 | Fuzz: five nightly corpus targets | 60 min build; 60 min per target | Gate; corpus/crash artifacts, scheduled failure alert | `MAX_TOTAL_TIME=<seconds> ./scripts/fuzz/run.sh` |
+| Daily 03:17 | MinIO interop (EC + SSE read parity) | 40 min | Gate; scheduled failure alert | Dispatch `minio-interop.yml` or follow its pinned Docker fixture steps |
+| Daily 04:29 | Replication / cluster-fault / protocol e2e | 45 / 90 / 90 min | Three independent gates; JUnit, membership, and server logs | `cargo nextest run --profile e2e-repl-nightly -p e2e_test`; `--profile e2e-nightly`; `-j 1 --profile e2e-protocols` |
+| Daily 06:31 | Warp performance A/B | 180 min | Regression budget gate; A/B summaries and server logs | `bash scripts/run_hotpath_warp_abba.sh --help` |
+| Daily 00:07 Asia/Shanghai (16:07 UTC previous day) | Nightly GNU build and Vault lanes | 150 / 90 / 60 min | Build, live Vault, and HA failover gates | Use the commands and pinned Vault images in `nightly-gnu.yml` |
+| Daily 03:23 | Security Audit | 20 / 5 min, plus 30 min on PR dependency review | Cargo Deny and workflow-pin gates; scheduled failure alert | `cargo deny check`; `scripts/security/check_workflow_pins.sh` |
+| Daily 23:47 | Scheduled Validation Freshness | 10 min | Fails when a critical schedule was never created or is stale | Dispatch `scheduled-validation-freshness.yml` |
+| Sunday 00:11 | Full `Continuous Integration` matrix | Per-job budgets above | Weekly variant coverage, including dormant rio-v2 binary/e2e lanes | Dispatch `ci.yml` |
+| Sunday 01:13 | Seven-platform build matrix | 150 min per platform | Build/package integrity; scheduled failure alert | Dispatch `build.yml` with an exact platform set |
+| Sunday 02:19 | Ceph s3-tests full sweep: single and real four-node, four shards each | 180 min per shard | Compatibility gate; report, JUnit, exact node IDs, and server logs | `scripts/s3-tests/run.sh` against an existing single or distributed target |
+| Sunday 06:41 | Mint | 120 min | **Report-only by design**; per-suite PASS/FAIL/NA and raw `log.json` | Reproduce the pinned Docker sequence in `mint.yml` or dispatch it |
+| Sunday 07:43 | Workspace line coverage | 120 min | Report-only trend; lcov and JSON retained 90 days | `make coverage` |
+| Monthly, day 1 06:37 | Runner Hygiene | 15 min | Validates runner ephemerality; scheduled failure alert | Dispatch `runner-hygiene.yml` |
+
+Manual `workflow_dispatch` exists for the scheduled workflows above. Manual
+runs are debugging evidence and intentionally do not open scheduled-failure
+issues. A manual performance run may explicitly allow a known regression; that
+override must not be treated as an ordinary passing baseline.
+
+## Release validation
+
+Release validation is post-merge and tag-driven; it does not substitute for a
+pull-request gate.
+
+| Event | Validation | Budget | Result |
+|---|---|---:|---|
+| Push to `main` or weekly schedule | `Build and Release` platform matrix | 150 min per platform | Build artifacts for all selected targets; no release publication on a main push |
+| Valid release or preview tag | `Build and Release` plus asset checks | 150 min per platform | Draft release, checksummed assets, and publish step |
+| Successful non-preview release-tag build | Docker image build and image scan | 60 min build; 30 min scan | Multi-architecture images plus vulnerability report |
+| Successful release-tag build | DEB/RPM packaging | 30 min per architecture | Packages and checksum files uploaded to the release |
+| Successful non-preview release-tag build | Helm template test and package | 30 min build; 30 min publish | Versioned chart and repository index |
+
+Use an exact preview tag for end-to-end release rehearsal. Manual dispatches
+are backfill/debug paths and do not prove the automatic `workflow_run` chain.
+
+## Evidence requirements
+
+A green check is useful only when it proves the intended behavior ran:
+
+- Record the exact commit SHA and run URL.
+- Separate product failure from runner prerequisites, service readiness, and
+  cancellation. Repair the precondition, then rerun the exact workload.
+- Preserve membership manifests, JUnit, raw compatibility logs, seeds, and
+  server logs where the workflow provides them.
+- For a bug fix or a new fault checker, provide sensitivity evidence: the old
+  behavior or an intentional mutation must fail the new oracle, and the fixed
+  behavior must pass it.
+- Never promote a report-only lane to required from one green run. Require at
+  least 14 days and 30 representative pull requests with at least 99% complete
+  execution, then update the ruleset and this table together.
+
+## Change checklist
+
+Update this file in the same pull request when any of these change:
+
+- workflow triggers, job names, timeouts, or nextest profile ownership;
+- required status contexts or strict/merge-queue policy;
+- scheduled cadence, alert routing, artifact contract, or local reproduction;
+- report-only versus gating semantics.
+
+Do not copy per-module test counts here. Update
+[e2e-suite-inventory.md](e2e-suite-inventory.md) and its enforced membership
+digest instead.

rustfs/Cargo.toml

@@ -322,7 +322,7 @@ thiserror = { workspace = true }
 tracing.workspace = true
 url = { workspace = true }
 urlencoding = { workspace = true }
-uuid = { workspace = true, features = ["v4", "fast-rng", "macro-diagnostics"] }
+uuid = { workspace = true, features = ["v4", "v5", "fast-rng", "macro-diagnostics"] }
 zip = { workspace = true }
 libc = { workspace = true }
 rand = { workspace = true, features = ["serde"] }
@@ -336,16 +336,16 @@ opentelemetry = { workspace = true }
 tracing-opentelemetry = { workspace = true }
 # Data structures
 hashbrown = { workspace = true, features = ["serde", "rayon"] }
-mimalloc = { workspace = true }
+rustfs-mimalloc = { workspace = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 libsystemd.workspace = true
 
 [target.'cfg(not(target_os = "windows"))'.dependencies]
-libmimalloc-sys.workspace = true
+rustfs-mimalloc-sys.workspace = true
 
 [dev-dependencies]
-uuid = { workspace = true, features = ["v4", "fast-rng", "macro-diagnostics"] }
+uuid = { workspace = true, features = ["v4", "v5", "fast-rng", "macro-diagnostics"] }
 serial_test = { workspace = true }
 tempfile = { workspace = true }
 aws-config = { workspace = true }

rustfs/src/admin/handlers/heal.rs

@@ -20,7 +20,7 @@ use crate::admin::storage_api::bucket::utils::is_valid_object_prefix;
 use crate::server::ADMIN_PREFIX;
 use crate::server::RemoteAddr;
 use crate::storage::rpc::node_service::heal::{
-    HealControlCoordinator, NodeHealProgress, NodeHealStatusSnapshot, capture_node_heal_status, decode_node_heal_status,
+    HealControlCoordinator, NodeHealStatusSnapshot, capture_node_heal_status, decode_node_heal_status,
     decode_node_replacement_recovery_status, heal_control_coordinator, heal_topology_fingerprint,
 };
 use bytes::Bytes;
@@ -298,14 +298,7 @@ fn background_heal_runtime_state(
     }
 }
 
-#[derive(Debug, Serialize)]
-#[serde(rename_all = "camelCase")]
-struct BackgroundHealProgress {
-    objects_scanned: u64,
-    objects_healed: u64,
-    objects_failed: u64,
-    bytes_processed: u64,
-}
+type BackgroundHealProgress = rustfs_heal::HealProgress;
 
 #[derive(Debug)]
 struct ClusterHealStatusSnapshot {
@@ -344,17 +337,10 @@ fn add_operations(total: &mut rustfs_heal::HealOperationsSnapshot, next: rustfs_
     add_source_counts(&mut total.retrying_by_source, next.retrying_by_source);
 }
 
-fn add_progress(total: &mut BackgroundHealProgress, next: NodeHealProgress) {
-    total.objects_scanned = total.objects_scanned.saturating_add(next.objects_scanned);
-    total.objects_healed = total.objects_healed.saturating_add(next.objects_healed);
-    total.objects_failed = total.objects_failed.saturating_add(next.objects_failed);
-    total.bytes_processed = total.bytes_processed.saturating_add(next.bytes_processed);
-}
-
 fn aggregate_cluster_heal_status(snapshots: Vec<NodeHealStatusSnapshot>) -> ClusterHealStatusSnapshot {
     let mut info = BackgroundHealInfo::default();
     let mut operations = rustfs_heal::HealOperationsSnapshot::default();
-    let mut progress = None;
+    let mut progress = Vec::new();
     let mut any_services_enabled = false;
     let mut any_initialized = false;
 
@@ -371,18 +357,12 @@ fn aggregate_cluster_heal_status(snapshots: Vec<NodeHealStatusSnapshot>) -> Clus
         }
         add_operations(&mut operations, snapshot.operations);
         if let Some(next) = snapshot.progress {
-            add_progress(
-                progress.get_or_insert(BackgroundHealProgress {
-                    objects_scanned: 0,
-                    objects_healed: 0,
-                    objects_failed: 0,
-                    bytes_processed: 0,
-                }),
-                next,
-            );
+            progress.push(next);
         }
     }
 
+    let progress = rustfs_heal::aggregate_heal_progress(progress);
+
     let state = if operations.queue_length > 0 || operations.active_tasks > 0 || operations.retrying_tasks > 0 {
         HealRuntimeState::Active
     } else if any_initialized {
@@ -2201,10 +2181,19 @@ mod tests {
         };
 
         let progress = BackgroundHealProgress {
+            kind: rustfs_heal::heal::progress::HealProgressKind::ObjectSweep,
             objects_scanned: 7,
             objects_healed: 3,
             objects_failed: 1,
+            skipped_objects: 3,
+            objects_total_count: 10,
+            objects_total_size: 8192,
             bytes_processed: 4096,
+            progress_percentage: 50.0,
+            progress_state: rustfs_heal::heal::progress::HealProgressState::Running,
+            baseline_generation: Some(42),
+            baseline_known: true,
+            ..Default::default()
         };
 
         let encoded = encode_background_heal_status(
@@ -2220,7 +2209,14 @@ mod tests {
         assert_eq!(json["progress"]["objectsScanned"], 7);
         assert_eq!(json["progress"]["objectsHealed"], 3);
         assert_eq!(json["progress"]["objectsFailed"], 1);
+        assert_eq!(json["progress"]["skippedObjects"], 3);
+        assert_eq!(json["progress"]["objectsTotalCount"], 10);
+        assert_eq!(json["progress"]["objectsTotalSize"], 8192);
         assert_eq!(json["progress"]["bytesProcessed"], 4096);
+        assert_eq!(json["progress"]["progressState"], "running");
+        assert_eq!(json["progress"]["baselineGeneration"], 42);
+        assert_eq!(json["progress"]["baselineKnown"], true);
+        assert_eq!(json["progress"]["counterUnknown"], false);
     }
 
     #[test]
@@ -2242,10 +2238,18 @@ mod tests {
                 ..Default::default()
             },
             Some(NodeHealProgress {
+                kind: rustfs_heal::heal::progress::HealProgressKind::ObjectSweep,
                 objects_scanned: 3,
                 objects_healed: 1,
                 objects_failed: 0,
+                skipped_objects: 2,
+                objects_total_count: 6,
+                objects_total_size: 400,
                 bytes_processed: 100,
+                progress_state: rustfs_heal::heal::progress::HealProgressState::Running,
+                baseline_generation: Some(9),
+                baseline_known: true,
+                ..Default::default()
             }),
         );
         let peer = NodeHealStatusSnapshot::for_test(
@@ -2265,10 +2269,17 @@ mod tests {
                 ..Default::default()
             },
             Some(NodeHealProgress {
+                kind: rustfs_heal::heal::progress::HealProgressKind::ObjectSweep,
                 objects_scanned: 5,
                 objects_healed: 4,
                 objects_failed: 1,
+                objects_total_count: 4,
+                objects_total_size: 1600,
                 bytes_processed: 900,
+                progress_state: rustfs_heal::heal::progress::HealProgressState::Running,
+                baseline_generation: Some(9),
+                baseline_known: true,
+                ..Default::default()
             }),
         );
 
@@ -2287,7 +2298,15 @@ mod tests {
         assert_eq!(progress.objects_scanned, 8);
         assert_eq!(progress.objects_healed, 5);
         assert_eq!(progress.objects_failed, 1);
+        assert_eq!(progress.skipped_objects, 2);
+        assert_eq!(progress.objects_total_count, 10);
+        assert_eq!(progress.objects_total_size, 2000);
         assert_eq!(progress.bytes_processed, 1000);
+        assert_eq!(progress.progress_percentage, 50.0);
+        assert_eq!(progress.progress_state, rustfs_heal::heal::progress::HealProgressState::Running);
+        assert_eq!(progress.baseline_generation, Some(9));
+        assert!(progress.baseline_known);
+        assert!(!progress.counter_unknown);
 
         assert_eq!(peer_first.state, HealRuntimeState::Active);
         assert_eq!(peer_first.operations, local_first.operations);
@@ -2327,6 +2346,7 @@ mod tests {
             objects_healed: value,
             objects_failed: value,
             bytes_processed: value,
+            ..Default::default()
         };
         let saturated = NodeHealStatusSnapshot::for_test(
             true,

rustfs/src/admin/handlers/site_replication.rs

@@ -41,7 +41,7 @@ use crate::admin::storage_api::config::save_admin_config;
 use crate::admin::storage_api::contract::bucket::{
     BucketOperations, BucketOptions, DeleteBucketOptions, MakeBucketOptions, SRBucketDeleteOp,
 };
-use crate::admin::storage_api::error::Error as StorageError;
+use crate::admin::storage_api::error::{Error as StorageError, is_err_bucket_not_found};
 use crate::admin::storage_api::runtime::ECStore;
 use crate::admin::utils::{encode_compatible_admin_payload, read_compatible_admin_body};
 use crate::auth::constant_time_eq;
@@ -55,6 +55,7 @@ use crate::storage::storage_api::{
 use base64::Engine;
 use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
 use base64::engine::general_purpose::URL_SAFE_NO_PAD;
+use futures::StreamExt;
 use hmac::{Hmac, Mac};
 use http::header::{CONTENT_TYPE, HOST};
 use http::{HeaderMap, HeaderValue, Uri};
@@ -2096,6 +2097,18 @@ async fn remote_add_preflight_info(site: &PeerSite) -> S3Result<SiteReplicationA
             format!("invalid site replication metainfo from `{}`: {e}", site.endpoint),
         )
     })?;
+    if info.deployment_id.is_empty() {
+        // The peer will be tracked under a locally derived fallback ID
+        // (deployment_id_for_endpoint) instead of its real deployment ID.
+        warn!(
+            event = EVENT_ADMIN_SITE_REPLICATION_STATE,
+            component = LOG_COMPONENT_ADMIN,
+            subsystem = LOG_SUBSYSTEM_SITE_REPLICATION,
+            result = "peer_deployment_id_missing",
+            peer_endpoint = %site.endpoint,
+            "admin site replication state"
+        );
+    }
 
     let idp_body = send_peer_admin_get_request_with_client(
         &client,
@@ -2206,20 +2219,30 @@ fn site_replication_bootstrap_token(uri: &Uri) -> Option<String> {
     query_pairs(uri).get("bootstrapToken").cloned()
 }
 
-fn bootstrap_bucket_make_op_path(bucket: &SRBucketInfo) -> String {
+/// Query for a peer `make-with-versioning` bucket op. `versioningEnabled`
+/// always travels so the outbound query matches MinIO's site-replication
+/// make-bucket wire contract: MinIO's own create-bucket hook sends
+/// `versioningEnabled=true` on this op. RustFS's inbound handler
+/// force-enables versioning either way.
+fn make_with_versioning_bucket_op_path(bucket: &str, created_at: Option<&str>, lock_enabled: bool) -> String {
     let mut query = form_urlencoded::Serializer::new(String::new());
-    query.append_pair("bucket", &bucket.bucket);
-    query.append_pair("operation", "make-with-versioning");
-    if let Some(created_at) = bucket
-        .created_at
-        .and_then(|value| value.format(&time::format_description::well_known::Rfc3339).ok())
-    {
-        query.append_pair("createdAt", &created_at);
+    query.append_pair("bucket", bucket);
+    query.append_pair("operation", SITE_REPLICATION_BUCKET_OP_MAKE_WITH_VERSIONING);
+    query.append_pair("versioningEnabled", "true");
+    if let Some(created_at) = created_at {
+        query.append_pair("createdAt", created_at);
     }
-    if bucket.object_lock_config.is_some() {
+    if lock_enabled {
         query.append_pair("lockEnabled", "true");
     }
-    format!("/rustfs/admin/v3/site-replication/peer/bucket-ops?{}", query.finish())
+    format!("{SITE_REPLICATION_PEER_BUCKET_OPS_PATH}?{}", query.finish())
+}
+
+fn bootstrap_bucket_make_op_path(bucket: &SRBucketInfo) -> String {
+    let created_at = bucket
+        .created_at
+        .and_then(|value| value.format(&time::format_description::well_known::Rfc3339).ok());
+    make_with_versioning_bucket_op_path(&bucket.bucket, created_at.as_deref(), bucket.object_lock_config.is_some())
 }
 
 fn bootstrap_bucket_meta_item(bucket: &SRBucketInfo, item_type: &str, updated_at: Option<OffsetDateTime>) -> SRBucketMeta {
@@ -4246,16 +4269,7 @@ async fn broadcast_site_replication_make_bucket(
         .format(&time::format_description::well_known::Rfc3339)
         .unwrap_or_default();
 
-    let path = {
-        let mut query = form_urlencoded::Serializer::new(String::new());
-        query.append_pair("bucket", bucket);
-        query.append_pair("operation", "make-with-versioning");
-        query.append_pair("createdAt", &created_at);
-        if lock_enabled {
-            query.append_pair("lockEnabled", "true");
-        }
-        format!("/rustfs/admin/v3/site-replication/peer/bucket-ops?{}", query.finish())
-    };
+    let path = make_with_versioning_bucket_op_path(bucket, Some(&created_at), lock_enabled);
     let path = if let Some(token) = bootstrap_token {
         with_site_replication_bootstrap_token(&path, token)
     } else {
@@ -10206,13 +10220,25 @@ impl Operation for SiteReplicationStatusHandler {
     }
 }
 
+/// `POST /v3/site-replication/devnull` — peer link-check upload drain.
+/// MinIO streams multi-megabyte probe bodies here during site netperf link
+/// checks and expects an unbounded discard (its handler copies to io.Discard);
+/// buffering through the 1MB admin body cap turned any larger probe into a
+/// 400 and a false link failure. Stream and discard instead — no size cap.
+async fn drain_site_replication_devnull(mut input: Body) -> S3Result<()> {
+    while let Some(chunk) = input.next().await {
+        chunk.map_err(|e| s3_error!(InvalidRequest, "failed to read devnull stream: {}", e))?;
+    }
+    Ok(())
+}
+
 pub struct SiteReplicationDevNullHandler {}
 
 #[async_trait::async_trait]
 impl Operation for SiteReplicationDevNullHandler {
     async fn call(&self, req: S3Request<Body>, _params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
         validate_site_replication_admin_request(&req, AdminAction::SiteReplicationOperationAction).await?;
-        let _ = read_plain_admin_body(req.input).await?;
+        drain_site_replication_devnull(req.input).await?;
         Ok(empty_response(StatusCode::NO_CONTENT))
     }
 }
@@ -10471,6 +10497,19 @@ impl Operation for SRPeerJoinHandler {
     }
 }
 
+/// Outcome of a peer-driven `purge-deleted-bucket` replay. A bucket that is
+/// already gone means the purge raced an earlier replay or a local delete —
+/// that is success — but any other failure must reach the sender like the
+/// sibling delete branches do: swallowing it answered 200 while the bucket
+/// survived on this site.
+fn purge_deleted_bucket_result(result: Result<(), StorageError>) -> S3Result<()> {
+    match result {
+        Ok(()) => Ok(()),
+        Err(err) if is_err_bucket_not_found(&err) => Ok(()),
+        Err(err) => Err(ApiError::from(err).into()),
+    }
+}
+
 pub struct SRPeerBucketOpsHandler {}
 
 #[async_trait::async_trait]
@@ -10570,16 +10609,18 @@ impl Operation for SRPeerBucketOpsHandler {
                     .map_err(ApiError::from)?;
             }
             "purge-deleted-bucket" => {
-                let _ = store
-                    .delete_bucket(
-                        &bucket,
-                        &DeleteBucketOptions {
-                            force: true,
-                            srdelete_op: SRBucketDeleteOp::Purge,
-                            ..Default::default()
-                        },
-                    )
-                    .await;
+                purge_deleted_bucket_result(
+                    store
+                        .delete_bucket(
+                            &bucket,
+                            &DeleteBucketOptions {
+                                force: true,
+                                srdelete_op: SRBucketDeleteOp::Purge,
+                                ..Default::default()
+                            },
+                        )
+                        .await,
+                )?;
             }
             _ => return Err(s3_error!(InvalidRequest, "unsupported site replication bucket operation")),
         }
@@ -13925,6 +13966,54 @@ mod tests {
         assert!(!query_flag(&uri, "missing"));
     }
 
+    /// A5 red-light: a `purge-deleted-bucket` replay must report success when
+    /// the bucket is already gone, and must propagate every other failure —
+    /// the swallowed error answered 200 while the bucket survived.
+    #[test]
+    fn test_purge_deleted_bucket_result_tolerates_only_missing_bucket() {
+        assert!(purge_deleted_bucket_result(Ok(())).is_ok());
+        assert!(purge_deleted_bucket_result(Err(StorageError::BucketNotFound("photos".to_string()))).is_ok());
+        assert!(purge_deleted_bucket_result(Err(StorageError::VolumeNotFound)).is_ok());
+        let err = purge_deleted_bucket_result(Err(StorageError::StorageFull))
+            .expect_err("non-not-found delete failures must propagate");
+        assert_ne!(*err.code(), S3ErrorCode::NoSuchBucket);
+    }
+
+    /// C5 red-light: the site-replication devnull drain must accept bodies
+    /// beyond the 1MB admin body cap — MinIO's link check streams large
+    /// probe bodies and treats a 400 as a broken link.
+    #[tokio::test]
+    async fn test_site_replication_devnull_drains_body_beyond_admin_cap() {
+        let body = Body::from(vec![0u8; MAX_ADMIN_REQUEST_BODY_SIZE + 1]);
+        drain_site_replication_devnull(body)
+            .await
+            .expect("devnull must drain bodies larger than the admin body cap");
+    }
+
+    /// A3 red-light: `versioningEnabled` must travel on every outbound
+    /// make-with-versioning bucket op so the query matches MinIO's
+    /// site-replication make-bucket wire contract (MinIO's own hook sends
+    /// `versioningEnabled=true` on this op).
+    #[test]
+    fn test_make_with_versioning_op_paths_send_versioning_enabled() {
+        let bucket = SRBucketInfo {
+            bucket: "photos".to_string(),
+            created_at: Some(OffsetDateTime::UNIX_EPOCH),
+            object_lock_config: Some(BASE64_STANDARD.encode("<ObjectLockConfiguration/>")),
+            ..Default::default()
+        };
+        let bootstrap = bootstrap_bucket_make_op_path(&bucket);
+        assert!(bootstrap.contains("operation=make-with-versioning"), "{bootstrap}");
+        assert!(bootstrap.contains("versioningEnabled=true"), "{bootstrap}");
+        assert!(bootstrap.contains("createdAt="), "{bootstrap}");
+        assert!(bootstrap.contains("lockEnabled=true"), "{bootstrap}");
+
+        // The broadcast path (create-bucket hook) shares the same builder.
+        let broadcast = make_with_versioning_bucket_op_path("photos", Some("1970-01-01T00:00:00Z"), false);
+        assert!(broadcast.contains("versioningEnabled=true"), "{broadcast}");
+        assert!(!broadcast.contains("lockEnabled"), "{broadcast}");
+    }
+
     #[tokio::test]
     #[serial]
     async fn test_add_bootstrap_scope_only_allows_expected_bucket_setup_until_guard_drops() {

rustfs/src/admin/site_replication_identity.rs

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 use rustfs_madmin::{PeerInfo, SyncStatus};
-use std::collections::{BTreeMap, hash_map::DefaultHasher};
-use std::hash::{Hash, Hasher};
+use std::collections::BTreeMap;
 use url::Url;
+use uuid::Uuid;
 
 fn has_http_scheme(endpoint: &str) -> bool {
     endpoint.get(..7).is_some_and(|prefix| prefix.eq_ignore_ascii_case("http://"))
@@ -66,10 +66,12 @@ pub fn site_identity_key(endpoint: &str) -> String {
         .unwrap_or_else(|| trimmed.to_ascii_lowercase())
 }
 
+/// Fallback deployment ID for a peer that reported none. UUIDv5 over the
+/// canonical endpoint: the ID is persisted in site-replication state and
+/// broadcast to peers, so it must be identical across Rust toolchains
+/// (`DefaultHasher` is not) and across spellings of the same endpoint.
 pub fn deployment_id_for_endpoint(endpoint: &str) -> String {
-    let mut hasher = DefaultHasher::new();
-    endpoint.hash(&mut hasher);
-    format!("{:016x}", hasher.finish())
+    Uuid::new_v5(&Uuid::NAMESPACE_URL, canonical_endpoint(endpoint).as_bytes()).to_string()
 }
 
 pub fn same_identity_endpoint(left: &str, right: &str) -> bool {
@@ -174,6 +176,23 @@ mod tests {
         }
     }
 
+    /// B8 red-light: the fallback deployment ID must be a toolchain-stable
+    /// UUIDv5 over the canonical endpoint — `DefaultHasher` output is not
+    /// guaranteed stable across Rust releases, yet the ID is persisted in
+    /// site-replication state and broadcast to peers.
+    #[test]
+    fn deployment_id_for_endpoint_is_stable_uuid_v5_over_canonical_endpoint() {
+        let endpoint = "https://node-a.example.com:9000";
+        let id = deployment_id_for_endpoint(endpoint);
+        let parsed = uuid::Uuid::parse_str(&id).expect("fallback deployment ID must be a UUID");
+        assert_eq!(parsed.get_version_num(), 5, "fallback deployment ID must be UUIDv5");
+        // Deterministic for the same endpoint and for spelling variants that
+        // share a canonical form; distinct endpoints stay distinct.
+        assert_eq!(id, deployment_id_for_endpoint(endpoint));
+        assert_eq!(id, deployment_id_for_endpoint(" HTTPS://Node-A.Example.Com:9000/ "));
+        assert_ne!(id, deployment_id_for_endpoint("https://node-b.example.com:9000"));
+    }
+
     #[test]
     fn canonical_endpoint_accepts_case_insensitive_scheme() {
         assert_eq!(

rustfs/src/admin/storage_api.rs

@@ -51,7 +51,7 @@ mod ecstore_disk {
 }
 
 mod ecstore_error {
-    pub(crate) use crate::storage::storage_api::ecstore_error::StorageError;
+    pub(crate) use crate::storage::storage_api::ecstore_error::{StorageError, is_err_bucket_not_found};
 }
 
 #[allow(unused_imports)]
@@ -919,6 +919,7 @@ pub(crate) mod contract {
 }
 
 pub(crate) mod error {
+    pub(crate) use super::ecstore_error::is_err_bucket_not_found;
     pub(crate) use super::{Error, StorageError};
 }
 

rustfs/src/allocator_reclaim.rs

@@ -369,14 +369,8 @@ pub fn allocator_reclaim_controller_snapshot(ctx: &CancellationToken) -> Allocat
 }
 
 #[cfg(not(target_os = "windows"))]
-#[allow(unsafe_code)]
 fn collect_allocator_memory(force: bool) -> Result<(), String> {
-    // SAFETY: `mi_collect` is provided by the active global allocator backend
-    // on this target family. It is explicitly intended to reclaim retained
-    // pages/segments and does not require additional invariants from the caller.
-    unsafe {
-        libmimalloc_sys::mi_collect(force);
-    }
+    rustfs_mimalloc::MiMalloc::collect(force);
     Ok(())
 }
 

rustfs/src/main.rs

@@ -26,22 +26,22 @@ struct MiMallocAllocator;
 unsafe impl GlobalAlloc for MiMallocAllocator {
     unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
         // SAFETY: the caller upholds GlobalAlloc's contract for layout.
-        unsafe { mimalloc::MiMalloc.alloc(layout) }
+        unsafe { rustfs_mimalloc::MiMalloc.alloc(layout) }
     }
 
     unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
         // SAFETY: the caller upholds GlobalAlloc's contract for layout.
-        unsafe { mimalloc::MiMalloc.alloc_zeroed(layout) }
+        unsafe { rustfs_mimalloc::MiMalloc.alloc_zeroed(layout) }
     }
 
     unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
         // SAFETY: ptr and layout came from this allocator and are forwarded unchanged.
-        unsafe { mimalloc::MiMalloc.dealloc(ptr, layout) }
+        unsafe { rustfs_mimalloc::MiMalloc.dealloc(ptr, layout) }
     }
 
     unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
         // SAFETY: ptr and layout came from this allocator and are forwarded unchanged.
-        unsafe { mimalloc::MiMalloc.realloc(ptr, layout, new_size) }
+        unsafe { rustfs_mimalloc::MiMalloc.realloc(ptr, layout, new_size) }
     }
 }
 
@@ -51,7 +51,7 @@ static GLOBAL: hotpath::CountingAllocator<MiMallocAllocator> = hotpath::Counting
 
 #[cfg(not(all(feature = "hotpath", feature = "hotpath-alloc")))]
 #[global_allocator]
-static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+static GLOBAL: rustfs_mimalloc::MiMalloc = rustfs_mimalloc::MiMalloc;
 
 fn main() {
     let _hotpath_guard = hotpath::HotpathGuardBuilder::new("main").build();
@@ -71,8 +71,9 @@ mod tests {
         allocation.extend_from_slice(&[7_u8; 64]);
 
         assert_eq!(allocation.len(), 64);
+        let heap = rustfs_mimalloc::heap::Heap::main();
         // SAFETY: the live Vec pointer is valid to inspect for heap ownership.
-        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(allocation.as_ptr().cast()) });
+        assert!(unsafe { heap.contains(allocation.as_ptr()) });
     }
 
     #[test]
@@ -85,12 +86,13 @@ mod tests {
         let layout = Layout::from_size_align(32, 8).expect("valid test allocation layout");
         let grown_layout = Layout::from_size_align(64, 8).expect("valid grown test allocation layout");
         let allocator = super::MiMallocAllocator;
+        let heap = rustfs_mimalloc::heap::Heap::main();
 
         // SAFETY: The pointer is checked for null before use and later released
         // through the same allocator with the corresponding layout.
         let ptr = unsafe { allocator.alloc_zeroed(layout) };
         assert!(!ptr.is_null());
-        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(ptr.cast()) });
+        assert!(unsafe { heap.contains(ptr) });
         assert!(unsafe { std::slice::from_raw_parts(ptr, 32).iter().all(|byte| *byte == 0) });
 
         // SAFETY: `ptr` was allocated by `allocator` with `layout`; on failure
@@ -102,7 +104,7 @@ mod tests {
             panic!("mimalloc realloc failed in allocator smoke test");
         }
 
-        assert!(unsafe { libmimalloc_sys::mi_is_in_heap_region(grown_ptr.cast()) });
+        assert!(unsafe { heap.contains(grown_ptr) });
         // SAFETY: `grown_ptr` was reallocated by `allocator` and is released
         // with the matching grown layout.
         unsafe { allocator.dealloc(grown_ptr, grown_layout) };

rustfs/src/memory_observability.rs

@@ -17,10 +17,7 @@ use rustfs_io_metrics::{
     record_cpu_usage, record_memory_usage, record_process_memory_split,
 };
 use serde::Serialize;
-#[cfg(any(test, not(target_os = "windows")))]
 use serde_json::Value;
-#[cfg(not(target_os = "windows"))]
-use std::ffi::CStr;
 use std::path::Path;
 use std::sync::{Arc, Mutex, OnceLock};
 use std::time::Duration;
@@ -231,7 +228,18 @@ fn read_cgroup_memory_snapshot() -> Option<CgroupMemorySnapshot> {
     read_cgroup_v2().or_else(read_cgroup_v1)
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
+fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
+    let json = rustfs_mimalloc::MiMalloc::stats_json();
+    if json.is_empty() {
+        return None;
+    }
+    let observation = parse_mimalloc_stats_json(&json)?;
+    Some(AllocatorMemorySnapshot {
+        backend: crate::allocator_reclaim::allocator_backend(),
+        observation,
+    })
+}
+
 fn numeric_json_value(value: &Value) -> Option<u64> {
     match value {
         Value::Number(number) => number
@@ -242,7 +250,6 @@ fn numeric_json_value(value: &Value) -> Option<u64> {
     }
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
 fn numeric_json_field(value: &Value, field: &str) -> Option<u64> {
     match value {
         Value::Object(fields) => fields
@@ -254,7 +261,6 @@ fn numeric_json_field(value: &Value, field: &str) -> Option<u64> {
     }
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_field(value: &Value, metric: &str, field: &str) -> Option<u64> {
     match value {
         Value::Object(fields) => {
@@ -271,12 +277,10 @@ fn mimalloc_stat_field(value: &Value, metric: &str, field: &str) -> Option<u64>
     }
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_current(value: &Value, metric: &str) -> Option<u64> {
     mimalloc_stat_field(value, metric, "current")
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
 fn mimalloc_stat_sum(value: &Value, metrics: &[&str], field: &str) -> Option<u64> {
     metrics
         .iter()
@@ -285,7 +289,6 @@ fn mimalloc_stat_sum(value: &Value, metrics: &[&str], field: &str) -> Option<u64
         .filter(|value| *value > 0)
 }
 
-#[cfg(any(test, not(target_os = "windows")))]
 fn parse_mimalloc_stats_json(stats_json: &str) -> Option<AllocatorMemoryObservation> {
     let value = serde_json::from_str::<Value>(stats_json).ok()?;
     let malloc_metrics = ["malloc_normal", "malloc_huge"];
@@ -312,33 +315,6 @@ fn parse_mimalloc_stats_json(stats_json: &str) -> Option<AllocatorMemoryObservat
     }
 }
 
-#[cfg(not(target_os = "windows"))]
-#[allow(unsafe_code)]
-fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
-    // SAFETY: `mi_stats_get_json` returns a null-terminated JSON buffer owned by
-    // mimalloc when called with a null input buffer. The mimalloc API requires
-    // freeing that buffer with `mi_free`; parsing finishes before the buffer is freed.
-    let observation = unsafe {
-        let stats_ptr = libmimalloc_sys::mi_stats_get_json(0, std::ptr::null_mut());
-        if stats_ptr.is_null() {
-            return None;
-        }
-
-        let observation = CStr::from_ptr(stats_ptr).to_str().ok().and_then(parse_mimalloc_stats_json);
-        libmimalloc_sys::mi_free(stats_ptr.cast());
-        observation?
-    };
-    Some(AllocatorMemorySnapshot {
-        backend: crate::allocator_reclaim::allocator_backend(),
-        observation,
-    })
-}
-
-#[cfg(target_os = "windows")]
-fn read_allocator_memory_snapshot() -> Option<AllocatorMemorySnapshot> {
-    None
-}
-
 fn configured_memory_observability_interval_secs() -> u64 {
     rustfs_utils::get_env_u64(ENV_MEMORY_OBSERVABILITY_INTERVAL_SECS, DEFAULT_MEMORY_OBSERVABILITY_INTERVAL_SECS).max(1)
 }
@@ -566,6 +542,13 @@ mod tests {
         assert_eq!(parse_mimalloc_stats_json(r#"{ "allocator": "unknown" }"#), None);
     }
 
+    #[test]
+    fn read_allocator_memory_snapshot_uses_mimalloc_stats_json() {
+        let snapshot = super::read_allocator_memory_snapshot();
+        #[cfg(not(target_os = "windows"))]
+        assert!(snapshot.is_some(), "allocator snapshot should be available on non-Windows");
+    }
+
     #[test]
     fn memory_observability_snapshot_reports_disabled_when_metrics_are_disabled() {
         let snapshot = build_memory_observability_status_snapshot(false, 15, false);

rustfs/src/storage/rpc/node_service.rs

@@ -1912,7 +1912,7 @@ impl Node for NodeService {
 
     async fn background_heal_status(
         &self,
-        _request: Request<BackgroundHealStatusRequest>,
+        request: Request<BackgroundHealStatusRequest>,
     ) -> Result<Response<BackgroundHealStatusResponse>, Status> {
         if self.resolve_object_store().is_none() {
             return Ok(Response::new(BackgroundHealStatusResponse {
@@ -1922,7 +1922,7 @@ impl Node for NodeService {
             }));
         }
         let snapshot = heal::capture_node_heal_status(rustfs_scanner::scanner::BackgroundHealInfo::default()).await;
-        match heal::encode_node_heal_status(&snapshot) {
+        match heal::encode_node_heal_status(&snapshot, request.into_inner().protocol_version) {
             Ok(bg_heal_state) => Ok(Response::new(BackgroundHealStatusResponse {
                 success: true,
                 bg_heal_state: bg_heal_state.into(),

rustfs/src/storage/rpc/node_service/heal.rs

@@ -25,7 +25,8 @@ use std::io::Cursor;
 
 use super::super::encode_msgpack_map;
 
-const NODE_HEAL_STATUS_VERSION: u8 = 1;
+const NODE_HEAL_STATUS_PREVIOUS_VERSION: u8 = 1;
+const NODE_HEAL_STATUS_VERSION: u8 = 2;
 const NODE_HEAL_STATUS_MAX_SIZE: usize = 64 * 1024;
 const NODE_REPLACEMENT_RECOVERY_STATUS_VERSION: u8 = 1;
 const NODE_REPLACEMENT_RECOVERY_STATUS_MAX_SIZE: usize = 64 * 1024;
@@ -172,13 +173,38 @@ pub(crate) fn heal_topology_fingerprint(endpoint_pools: &EndpointServerPools) ->
     Ok(hex_simd::encode_to_string(hasher.finalize(), hex_simd::AsciiCase::Lower))
 }
 
-#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) type NodeHealProgress = rustfs_heal::HealProgress;
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase", deny_unknown_fields)]
-pub(crate) struct NodeHealProgress {
-    pub objects_scanned: u64,
-    pub objects_healed: u64,
-    pub objects_failed: u64,
-    pub bytes_processed: u64,
+struct NodeHealProgressV1 {
+    objects_scanned: u64,
+    objects_healed: u64,
+    objects_failed: u64,
+    bytes_processed: u64,
+}
+
+impl From<&NodeHealProgress> for NodeHealProgressV1 {
+    fn from(progress: &NodeHealProgress) -> Self {
+        Self {
+            objects_scanned: progress.objects_scanned,
+            objects_healed: progress.objects_healed,
+            objects_failed: progress.objects_failed,
+            bytes_processed: progress.bytes_processed,
+        }
+    }
+}
+
+impl From<NodeHealProgressV1> for NodeHealProgress {
+    fn from(progress: NodeHealProgressV1) -> Self {
+        Self {
+            objects_scanned: progress.objects_scanned,
+            objects_healed: progress.objects_healed,
+            objects_failed: progress.objects_failed,
+            bytes_processed: progress.bytes_processed,
+            ..Default::default()
+        }
+    }
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -220,6 +246,48 @@ pub(crate) struct NodeHealStatusSnapshot {
     pub progress: Option<NodeHealProgress>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase", deny_unknown_fields)]
+struct NodeHealStatusSnapshotV1 {
+    version: u8,
+    services_enabled: bool,
+    initialized: bool,
+    info: NodeHealInfo,
+    operations: HealOperationsSnapshot,
+    progress: Option<NodeHealProgressV1>,
+}
+
+impl From<&NodeHealStatusSnapshot> for NodeHealStatusSnapshotV1 {
+    fn from(snapshot: &NodeHealStatusSnapshot) -> Self {
+        Self {
+            version: NODE_HEAL_STATUS_PREVIOUS_VERSION,
+            services_enabled: snapshot.services_enabled,
+            initialized: snapshot.initialized,
+            info: snapshot.info.clone(),
+            operations: snapshot.operations,
+            progress: snapshot.progress.as_ref().map(NodeHealProgressV1::from),
+        }
+    }
+}
+
+impl From<NodeHealStatusSnapshotV1> for NodeHealStatusSnapshot {
+    fn from(snapshot: NodeHealStatusSnapshotV1) -> Self {
+        Self {
+            version: snapshot.version,
+            services_enabled: snapshot.services_enabled,
+            initialized: snapshot.initialized,
+            info: snapshot.info,
+            operations: snapshot.operations,
+            progress: snapshot.progress.map(NodeHealProgress::from),
+        }
+    }
+}
+
+#[derive(Deserialize)]
+struct NodeHealStatusVersion {
+    version: u8,
+}
+
 impl NodeHealStatusSnapshot {
     #[cfg(test)]
     pub(crate) fn for_test(
@@ -249,14 +317,7 @@ impl NodeHealStatusSnapshot {
 }
 
 pub(crate) async fn capture_node_heal_status(info: BackgroundHealInfo) -> NodeHealStatusSnapshot {
-    let progress = rustfs_heal::current_heal_progress_snapshot()
-        .await
-        .map(|progress| NodeHealProgress {
-            objects_scanned: progress.objects_scanned,
-            objects_healed: progress.objects_healed,
-            objects_failed: progress.objects_failed,
-            bytes_processed: progress.bytes_processed,
-        });
+    let progress = rustfs_heal::current_heal_progress_snapshot().await;
 
     NodeHealStatusSnapshot {
         version: NODE_HEAL_STATUS_VERSION,
@@ -268,22 +329,43 @@ pub(crate) async fn capture_node_heal_status(info: BackgroundHealInfo) -> NodeHe
     }
 }
 
-pub(crate) fn encode_node_heal_status(snapshot: &NodeHealStatusSnapshot) -> Result<Vec<u8>, String> {
-    encode_msgpack_map(snapshot).map_err(|err| format!("failed to encode node heal status: {err}"))
+pub(crate) fn encode_node_heal_status(snapshot: &NodeHealStatusSnapshot, protocol_version: u32) -> Result<Vec<u8>, String> {
+    let encoded = if protocol_version < rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION {
+        encode_msgpack_map(&NodeHealStatusSnapshotV1::from(snapshot))
+    } else {
+        let mut snapshot = snapshot.clone();
+        snapshot.version = NODE_HEAL_STATUS_VERSION;
+        encode_msgpack_map(&snapshot)
+    };
+    encoded.map_err(|err| format!("failed to encode node heal status: {err}"))
 }
 
 pub(crate) fn decode_node_heal_status(data: &[u8]) -> Result<NodeHealStatusSnapshot, String> {
     if data.len() > NODE_HEAL_STATUS_MAX_SIZE {
         return Err("node heal status exceeds size limit".to_string());
     }
-    let mut deserializer = Deserializer::new(Cursor::new(data));
-    let snapshot = NodeHealStatusSnapshot::deserialize(&mut deserializer)
-        .map_err(|err| format!("failed to decode node heal status: {err}"))?;
+    let decode_version = || {
+        let mut deserializer = Deserializer::new(Cursor::new(data));
+        NodeHealStatusVersion::deserialize(&mut deserializer)
+            .map(|version| (version, deserializer))
+            .map_err(|err| format!("failed to decode node heal status: {err}"))
+    };
+    let (version, deserializer) = decode_version()?;
     if usize::try_from(deserializer.get_ref().position()).ok() != Some(data.len()) {
         return Err("node heal status contains trailing data".to_string());
     }
-    if snapshot.version != NODE_HEAL_STATUS_VERSION {
-        return Err(format!("unsupported node heal status version: {}", snapshot.version));
+
+    let mut deserializer = Deserializer::new(Cursor::new(data));
+    let snapshot = match version.version {
+        NODE_HEAL_STATUS_PREVIOUS_VERSION => {
+            NodeHealStatusSnapshotV1::deserialize(&mut deserializer).map(NodeHealStatusSnapshot::from)
+        }
+        NODE_HEAL_STATUS_VERSION => NodeHealStatusSnapshot::deserialize(&mut deserializer),
+        version => return Err(format!("unsupported node heal status version: {version}")),
+    }
+    .map_err(|err| format!("failed to decode node heal status: {err}"))?;
+    if usize::try_from(deserializer.get_ref().position()).ok() != Some(data.len()) {
+        return Err("node heal status contains trailing data".to_string());
     }
     Ok(snapshot)
 }
@@ -387,9 +469,10 @@ pub(crate) fn decode_node_replacement_recovery_status(data: &[u8]) -> Result<Nod
 #[cfg(test)]
 mod tests {
     use super::{
-        NODE_HEAL_STATUS_MAX_SIZE, NODE_HEAL_STATUS_VERSION, NodeHealProgress, NodeHealStatusSnapshot,
-        NodeReplacementRecoveryStatusSnapshot, decode_node_heal_status, decode_node_replacement_recovery_status,
-        encode_node_heal_status, encode_node_replacement_recovery_status, heal_control_coordinator, heal_topology_fingerprint,
+        NODE_HEAL_STATUS_MAX_SIZE, NODE_HEAL_STATUS_PREVIOUS_VERSION, NODE_HEAL_STATUS_VERSION, NodeHealProgress,
+        NodeHealStatusSnapshot, NodeReplacementRecoveryStatusSnapshot, decode_node_heal_status,
+        decode_node_replacement_recovery_status, encode_node_heal_status, encode_node_replacement_recovery_status,
+        heal_control_coordinator, heal_topology_fingerprint,
     };
     use crate::storage::storage_api::{
         Endpoint,
@@ -533,20 +616,72 @@ mod tests {
                 ..Default::default()
             },
             Some(NodeHealProgress {
+                kind: rustfs_heal::heal::progress::HealProgressKind::ObjectSweep,
                 objects_scanned: 7,
                 objects_healed: 5,
                 objects_failed: 1,
+                skipped_objects: 1,
+                objects_total_count: 10,
+                objects_total_size: 2048,
                 bytes_processed: 1024,
+                progress_percentage: 50.0,
+                progress_state: rustfs_heal::heal::progress::HealProgressState::Running,
+                baseline_generation: Some(42),
+                baseline_known: true,
+                ..Default::default()
             }),
         );
 
-        let encoded = encode_node_heal_status(&snapshot).expect("snapshot should encode");
+        let encoded = encode_node_heal_status(&snapshot, rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION)
+            .expect("snapshot should encode");
         let decoded = decode_node_heal_status(&encoded).expect("snapshot should decode");
         assert_eq!(decoded.version, NODE_HEAL_STATUS_VERSION);
         assert_eq!(decoded.operations.queue_length, 2);
         assert_eq!(decoded.progress, snapshot.progress);
     }
 
+    #[test]
+    fn node_heal_status_v1_encoding_preserves_rolling_compatibility() {
+        let snapshot = NodeHealStatusSnapshot::for_test(
+            true,
+            true,
+            BackgroundHealInfo::default(),
+            HealOperationsSnapshot::default(),
+            Some(NodeHealProgress {
+                objects_scanned: 7,
+                objects_healed: 5,
+                objects_failed: 1,
+                skipped_objects: 3,
+                bytes_processed: 1024,
+                progress_state: rustfs_heal::heal::progress::HealProgressState::Running,
+                baseline_known: true,
+                ..Default::default()
+            }),
+        );
+
+        let encoded =
+            encode_node_heal_status(&snapshot, u32::from(NODE_HEAL_STATUS_PREVIOUS_VERSION)).expect("v1 snapshot should encode");
+        let wire: serde_json::Value = rmp_serde::from_slice(&encoded).expect("v1 snapshot should decode as JSON");
+        let progress = wire["progress"].as_object().expect("v1 progress should be a map");
+        assert_eq!(wire["version"], NODE_HEAL_STATUS_PREVIOUS_VERSION);
+        assert_eq!(progress.len(), 4);
+        assert_eq!(progress["objectsScanned"], 7);
+        assert!(!progress.contains_key("skippedObjects"));
+
+        let decoded = decode_node_heal_status(&encoded).expect("v1 snapshot should decode");
+        let progress = decoded.progress.as_ref().expect("v1 progress should be present");
+        assert_eq!(decoded.version, NODE_HEAL_STATUS_PREVIOUS_VERSION);
+        assert_eq!(progress.objects_scanned, 7);
+        assert_eq!(progress.progress_state, rustfs_heal::heal::progress::HealProgressState::Unknown);
+        assert!(!progress.baseline_known);
+
+        let upgraded = encode_node_heal_status(&decoded, rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION)
+            .expect("decoded v1 snapshot should upgrade to v2");
+        let upgraded = decode_node_heal_status(&upgraded).expect("upgraded snapshot should decode");
+        assert_eq!(upgraded.version, NODE_HEAL_STATUS_VERSION);
+        assert_eq!(upgraded.progress.as_ref(), Some(progress));
+    }
+
     #[test]
     fn node_heal_status_rejects_unknown_version() {
         let mut snapshot = NodeHealStatusSnapshot::for_test(
@@ -558,7 +693,7 @@ mod tests {
         );
         snapshot.version += 1;
 
-        let encoded = encode_node_heal_status(&snapshot).expect("snapshot should encode");
+        let encoded = rmp_serde::to_vec_named(&snapshot).expect("snapshot should encode");
         let err = decode_node_heal_status(&encoded).expect_err("unknown version should fail closed");
         assert!(err.contains("unsupported node heal status version"));
     }
@@ -579,13 +714,22 @@ mod tests {
                 "activeBySource": {"scanner": 0, "admin": 1, "autoHeal": 0, "internal": 0, "readRepair": 0},
                 "retryingBySource": {"scanner": 0, "admin": 0, "autoHeal": 0, "internal": 0, "readRepair": 0}
             },
-            "progress": null
+            "progress": {
+                "objectsScanned": 7,
+                "objectsHealed": 5,
+                "objectsFailed": 1,
+                "bytesProcessed": 1024
+            }
         });
         let encoded = rmp_serde::to_vec_named(&fixture).expect("fixture should encode");
         let decoded = decode_node_heal_status(&encoded).expect("fixed v1 fixture should decode");
         assert_eq!(decoded.info().bitrot_start_cycle, 9);
         assert_eq!(decoded.operations.queue_length, 2);
         assert_eq!(decoded.operations.queued_by_source.mrf, 0);
+        let progress = decoded.progress.expect("legacy progress should decode");
+        assert_eq!(progress.objects_scanned, 7);
+        assert!(!progress.baseline_known);
+        assert_eq!(progress.progress_state, rustfs_heal::heal::progress::HealProgressState::Unknown);
     }
 
     #[test]
@@ -631,7 +775,8 @@ mod tests {
             HealOperationsSnapshot::default(),
             None,
         );
-        let encoded = encode_node_heal_status(&snapshot).expect("snapshot should encode");
+        let encoded = encode_node_heal_status(&snapshot, rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION)
+            .expect("snapshot should encode");
         let encoded_json: serde_json::Value = rmp_serde::from_slice(&encoded).expect("encoded snapshot should decode as JSON");
         assert_eq!(encoded_json["info"]["bitrotStartTime"], serde_json::json!("2023-11-14T22:13:20.123456Z"));
     }