fix(heal): preserve resumable bucket checkpoints

crates/ecstore/src/core/pools.rs

@@ -865,22 +865,16 @@ fn should_replace_pool_status_for_status_refresh(
     !has_active_worker && persisted.last_update > current.last_update
 }
 
-/// Merges a persisted pool metadata snapshot into `current` monotonically:
-/// a pool entry is replaced only when no active worker covers it and the
-/// snapshot is strictly newer, so delayed snapshots never roll back local
-/// queued/terminal progressions. Returns whether any entry was replaced or
-/// appended.
-pub(crate) fn merge_pool_status_refresh(current: &mut PoolMeta, persisted: PoolMeta, active_workers: &[bool]) -> bool {
+fn merge_pool_status_refresh(current: &mut PoolMeta, persisted: PoolMeta, active_workers: &[bool]) {
     if persisted.pools.is_empty() {
-        return false;
+        return;
     }
 
     if current.pools.is_empty() {
         *current = persisted;
-        return true;
+        return;
     }
 
-    let mut merged_newer = false;
     for (idx, persisted_pool) in persisted.pools.into_iter().enumerate() {
         if persisted_pool.id != idx {
             continue;
@@ -890,14 +884,11 @@ pub(crate) fn merge_pool_status_refresh(current: &mut PoolMeta, persisted: PoolM
         if idx < current.pools.len() {
             if should_replace_pool_status_for_status_refresh(current.pools.get(idx), &persisted_pool, has_active_worker) {
                 current.pools[idx] = persisted_pool;
-                merged_newer = true;
             }
         } else if idx == current.pools.len() && !has_active_worker {
             current.pools.push(persisted_pool);
-            merged_newer = true;
         }
     }
-    merged_newer
 }
 
 fn resolve_start_decommission_pool_meta_reload_result(result: Result<()>) -> Result<()> {
@@ -5716,55 +5707,6 @@ mod pools_tests {
         assert_eq!(info.bytes_done, 1_024);
     }
 
-    #[test]
-    fn test_merge_pool_status_refresh_fails_closed_on_missing_persisted_pools() {
-        let newer = OffsetDateTime::from_unix_timestamp(2_000).expect("test timestamp should be valid");
-        let mut current = PoolMeta {
-            pools: vec![decommission_test_pool_status(
-                0,
-                Some(PoolDecommissionInfo {
-                    complete: true,
-                    ..Default::default()
-                }),
-            )],
-            ..Default::default()
-        };
-        current.pools[0].last_update = newer;
-
-        assert!(
-            !merge_pool_status_refresh(&mut current, PoolMeta::default(), &[false]),
-            "an empty persisted snapshot must fail closed instead of replacing local state"
-        );
-
-        let info = current.pools[0]
-            .decommission
-            .as_ref()
-            .expect("local decommission info should survive a missing snapshot");
-        assert!(info.complete);
-        assert_eq!(current.pools[0].last_update, newer);
-    }
-
-    #[test]
-    fn test_merge_pool_status_refresh_ignores_mislabeled_pool_entries() {
-        let older = OffsetDateTime::from_unix_timestamp(1_000).expect("test timestamp should be valid");
-        let mut current = PoolMeta {
-            pools: vec![decommission_test_pool_status(0, None)],
-            ..Default::default()
-        };
-        let mut persisted = PoolMeta {
-            pools: vec![decommission_test_pool_status(0, Some(PoolDecommissionInfo::default()))],
-            ..Default::default()
-        };
-        persisted.pools[0].id = 7;
-        persisted.pools[0].last_update = older;
-
-        assert!(
-            !merge_pool_status_refresh(&mut current, persisted, &[false]),
-            "a pool entry whose id does not match its index must be ignored"
-        );
-        assert!(current.pools[0].decommission.is_none());
-    }
-
     #[test]
     fn test_dedup_indices_removes_duplicates_preserving_order() {
         assert_eq!(dedup_indices(&[0, 2, 1, 2, 3, 0]), vec![0, 2, 1, 3]);

crates/ecstore/src/store/rebalance.rs

@@ -14,15 +14,10 @@
 
 use super::*;
 use crate::config::storageclass;
-use crate::core::pools::merge_pool_status_refresh;
 use crate::layout::pool_space::{ServerPoolsAvailableSpace, build_server_pools_available_space};
 use crate::runtime::sources as runtime_sources;
 use crate::storage_api_contracts::{admin::StorageAdminApi, namespace::NamespaceLocking as _, object::ObjectOperations as _};
 pub(in crate::store) mod support;
-
-const LOG_COMPONENT_ECSTORE: &str = "ecstore";
-const LOG_SUBSYSTEM_POOLS: &str = "pools";
-const EVENT_POOL_META_RELOAD: &str = "pool_meta_reload";
 use support::{
     LatestObjectInfoCandidate, PoolErr, PoolObjInfo, RebalanceDeletePoolResult, pool_lookup_not_found_error,
     rebalance_disk_set_lookup_error, resolve_latest_object_info_candidates, resolve_rebalance_delete_from_all_pools_result,
@@ -689,49 +684,17 @@ impl ECStore {
         )
     }
 
-    /// Peer reload entry: refreshes in-memory pool metadata from the shared
-    /// persisted snapshot. Returns whether newer state was actually merged so
-    /// callers only trigger missing-worker recovery after a real state change;
-    /// delayed snapshots are merged monotonically and never blind-assigned.
-    pub async fn reload_pool_meta(&self) -> Result<bool> {
-        let mut reloaded = PoolMeta::default();
+    pub async fn reload_pool_meta(&self) -> Result<()> {
+        let mut meta = PoolMeta::default();
         resolve_store_rebalance_pool_meta_reload_result(
-            reloaded.load(self.pools[0].clone(), self.pools.clone()).await,
+            meta.load(self.pools[0].clone(), self.pools.clone()).await,
             "reload_pool_meta",
         )?;
 
-        // Lock order: release the decommission_cancelers guard before taking
-        // the pool_meta write guard; neither is held across the disk read.
-        let active_workers = {
-            let cancelers = self.decommission_cancelers.read().await;
-            cancelers.iter().map(Option::is_some).collect::<Vec<_>>()
-        };
-
-        let incoming_has_pools = !reloaded.pools.is_empty();
         let mut pool_meta = self.pool_meta.write().await;
-        let merged_newer = merge_pool_status_refresh(&mut pool_meta, reloaded, &active_workers);
-
-        if !merged_newer && !incoming_has_pools {
-            warn!(
-                event = EVENT_POOL_META_RELOAD,
-                component = LOG_COMPONENT_ECSTORE,
-                subsystem = LOG_SUBSYSTEM_POOLS,
-                result = "ignored",
-                reason = "missing_metadata",
-                "Peer pool meta reload ignored because persisted metadata is missing"
-            );
-        } else if !merged_newer {
-            debug!(
-                event = EVENT_POOL_META_RELOAD,
-                component = LOG_COMPONENT_ECSTORE,
-                subsystem = LOG_SUBSYSTEM_POOLS,
-                result = "ignored",
-                reason = "stale_snapshot",
-                "Peer pool meta reload ignored as a stale snapshot"
-            );
-        }
-
-        Ok(merged_newer)
+        *pool_meta = meta;
+        // *self.pool_meta.write().expect("operation should succeed") = meta;
+        Ok(())
     }
 
     /// Disk information deduplication function
@@ -897,7 +860,6 @@ fn lifecycle_delete_all_test_failure(phase: crate::object_api::LifecycleDeleteAl
 mod tests {
     use super::*;
     use crate::config::storageclass::{CLASS_RRS, CLASS_STANDARD, lookup_config_for_pools_without_env};
-    use crate::core::pools::{POOL_META_VERSION, PoolDecommissionInfo, PoolStatus};
     use crate::disk::error::DiskError;
     use crate::layout::endpoint::Endpoint;
     use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
@@ -908,7 +870,6 @@ mod tests {
     use rustfs_config::server_config::KVS;
     use rustfs_filemeta::FileInfo;
     use std::sync::Arc;
-    use time::{Duration as TimeDuration, OffsetDateTime};
     use tokio_util::sync::CancellationToken;
 
     async fn setup_multi_pool_test_store(
@@ -1751,280 +1712,4 @@ mod tests {
                 .contains("failed to resolve rebalance disk set: pool index 2, set index 7, pool count 3")
         );
     }
-
-    fn reload_test_pool_status(decommission: Option<PoolDecommissionInfo>, last_update: time::OffsetDateTime) -> PoolStatus {
-        PoolStatus {
-            id: 0,
-            cmd_line: "pool-0".to_string(),
-            last_update,
-            decommission,
-        }
-    }
-
-    fn reload_test_pool_meta(pool: PoolStatus) -> PoolMeta {
-        PoolMeta {
-            version: POOL_META_VERSION,
-            pools: vec![pool],
-            dont_save: false,
-        }
-    }
-
-    async fn persist_reload_snapshot(store: &ECStore, snapshot: &PoolMeta) {
-        snapshot
-            .save(store.pools.clone())
-            .await
-            .expect("pool meta snapshot should persist to every pool");
-    }
-
-    #[tokio::test]
-    #[serial_test::serial]
-    async fn peer_pool_meta_reload_does_not_rollback_newer_local_states() {
-        let (_temp_dir, store, shutdown) = setup_multi_pool_test_store("pool-meta-reload-stale", &[2]).await;
-
-        let stale_time = OffsetDateTime::now_utc();
-        let newer_time = stale_time + TimeDuration::seconds(30);
-        let progressed_states: [(&str, PoolDecommissionInfo); 4] = [
-            (
-                "queued",
-                PoolDecommissionInfo {
-                    queued: true,
-                    start_time: Some(stale_time),
-                    ..Default::default()
-                },
-            ),
-            (
-                "canceled",
-                PoolDecommissionInfo {
-                    canceled: true,
-                    start_time: Some(stale_time),
-                    ..Default::default()
-                },
-            ),
-            (
-                "failed",
-                PoolDecommissionInfo {
-                    failed: true,
-                    start_time: Some(stale_time),
-                    ..Default::default()
-                },
-            ),
-            (
-                "complete",
-                PoolDecommissionInfo {
-                    complete: true,
-                    start_time: Some(stale_time),
-                    ..Default::default()
-                },
-            ),
-        ];
-
-        for (state_label, local_state) in progressed_states {
-            {
-                let mut pool_meta = store.pool_meta.write().await;
-                *pool_meta = reload_test_pool_meta(reload_test_pool_status(Some(local_state.clone()), newer_time));
-            }
-
-            // A delayed peer message carries a snapshot that predates the local progression.
-            let stale_snapshot = reload_test_pool_meta(reload_test_pool_status(
-                Some(PoolDecommissionInfo {
-                    start_time: Some(stale_time),
-                    ..Default::default()
-                }),
-                stale_time,
-            ));
-            persist_reload_snapshot(&store, &stale_snapshot).await;
-
-            let merged_newer = store.reload_pool_meta().await.expect("stale reload should succeed");
-            assert!(
-                !merged_newer,
-                "a delayed reload must not report merged newer state for the {state_label} progression"
-            );
-
-            let pool_meta = store.pool_meta.read().await;
-            let info = pool_meta.pools[0]
-                .decommission
-                .as_ref()
-                .expect("local decommission state should survive a stale reload");
-            assert_eq!(info.queued, local_state.queued, "{state_label} queued flag must not roll back");
-            assert_eq!(info.canceled, local_state.canceled, "{state_label} canceled flag must not roll back");
-            assert_eq!(info.failed, local_state.failed, "{state_label} failed flag must not roll back");
-            assert_eq!(info.complete, local_state.complete, "{state_label} complete flag must not roll back");
-            assert_eq!(
-                pool_meta.pools[0].last_update, newer_time,
-                "{state_label} progress timestamp must be kept"
-            );
-        }
-
-        shutdown.cancel();
-    }
-
-    #[tokio::test]
-    #[serial_test::serial]
-    async fn peer_pool_meta_reload_merges_newer_state_and_is_idempotent_on_duplicate_delivery() {
-        let (_temp_dir, store, shutdown) = setup_multi_pool_test_store("pool-meta-reload-duplicate", &[2]).await;
-
-        let older_time = OffsetDateTime::now_utc();
-        let newer_time = older_time + TimeDuration::seconds(30);
-
-        {
-            let mut pool_meta = store.pool_meta.write().await;
-            *pool_meta = reload_test_pool_meta(reload_test_pool_status(
-                Some(PoolDecommissionInfo {
-                    items_decommissioned: 1,
-                    ..Default::default()
-                }),
-                older_time,
-            ));
-        }
-        let newer_snapshot = reload_test_pool_meta(reload_test_pool_status(
-            Some(PoolDecommissionInfo {
-                complete: true,
-                items_decommissioned: 10,
-                ..Default::default()
-            }),
-            newer_time,
-        ));
-        persist_reload_snapshot(&store, &newer_snapshot).await;
-
-        let merged_newer = store.reload_pool_meta().await.expect("first reload should succeed");
-        assert!(merged_newer, "a strictly newer persisted snapshot must merge");
-
-        {
-            let pool_meta = store.pool_meta.read().await;
-            let info = pool_meta.pools[0].decommission.as_ref().expect("merged decommission state");
-            assert!(info.complete);
-            assert_eq!(info.items_decommissioned, 10);
-            assert_eq!(pool_meta.pools[0].last_update, newer_time);
-        }
-
-        // Redelivering the same generation must be a no-op.
-        let duplicate_merged = store.reload_pool_meta().await.expect("duplicate reload should succeed");
-        assert!(!duplicate_merged, "a duplicate delivery must not re-apply merged state");
-
-        {
-            let pool_meta = store.pool_meta.read().await;
-            let info = pool_meta.pools[0].decommission.as_ref().expect("merged decommission state");
-            assert!(info.complete);
-            assert_eq!(info.items_decommissioned, 10);
-            assert_eq!(pool_meta.pools[0].last_update, newer_time);
-        }
-
-        shutdown.cancel();
-    }
-
-    #[tokio::test]
-    #[serial_test::serial]
-    async fn peer_pool_meta_reload_keeps_active_worker_progress_over_newer_snapshot() {
-        let (_temp_dir, store, shutdown) = setup_multi_pool_test_store("pool-meta-reload-worker", &[2]).await;
-        *store.decommission_cancelers.write().await = vec![Some(CancellationToken::new())];
-
-        let worker_time = OffsetDateTime::now_utc();
-        let newer_time = worker_time + TimeDuration::seconds(30);
-
-        {
-            let mut pool_meta = store.pool_meta.write().await;
-            *pool_meta = reload_test_pool_meta(reload_test_pool_status(
-                Some(PoolDecommissionInfo {
-                    start_time: Some(worker_time),
-                    items_decommissioned: 10,
-                    bytes_done: 1_024,
-                    ..Default::default()
-                }),
-                worker_time,
-            ));
-        }
-        // Even a strictly newer terminal snapshot must not override a live worker.
-        let newer_terminal_snapshot = reload_test_pool_meta(reload_test_pool_status(
-            Some(PoolDecommissionInfo {
-                complete: true,
-                ..Default::default()
-            }),
-            newer_time,
-        ));
-        persist_reload_snapshot(&store, &newer_terminal_snapshot).await;
-
-        let merged_newer = store
-            .reload_pool_meta()
-            .await
-            .expect("reload under an active worker should succeed");
-        assert!(!merged_newer, "an active local worker must block snapshot replacement");
-
-        let pool_meta = store.pool_meta.read().await;
-        let info = pool_meta.pools[0]
-            .decommission
-            .as_ref()
-            .expect("worker progress should remain");
-        assert!(!info.complete);
-        assert_eq!(info.items_decommissioned, 10);
-        assert_eq!(info.bytes_done, 1_024);
-        assert_eq!(pool_meta.pools[0].last_update, worker_time);
-
-        shutdown.cancel();
-    }
-
-    #[tokio::test]
-    #[serial_test::serial]
-    async fn peer_pool_meta_reload_fails_closed_when_persisted_metadata_is_missing() {
-        let (temp_dir, store, shutdown) = setup_multi_pool_test_store("pool-meta-reload-missing", &[2]).await;
-
-        let kept_time = OffsetDateTime::now_utc();
-        {
-            let mut pool_meta = store.pool_meta.write().await;
-            *pool_meta = reload_test_pool_meta(reload_test_pool_status(
-                Some(PoolDecommissionInfo {
-                    complete: true,
-                    ..Default::default()
-                }),
-                kept_time,
-            ));
-        }
-        // Persist first so the test controls exactly what exists on disk.
-        persist_reload_snapshot(
-            &store,
-            &reload_test_pool_meta(reload_test_pool_status(
-                Some(PoolDecommissionInfo {
-                    complete: true,
-                    ..Default::default()
-                }),
-                kept_time,
-            )),
-        )
-        .await;
-
-        let mut deleted_any = false;
-        for disk_index in 0..2 {
-            let pool_bin_dir = temp_dir
-                .path()
-                .join(format!("pool0-disk{disk_index}"))
-                .join(crate::disk::RUSTFS_META_BUCKET)
-                .join(crate::core::pools::POOL_META_NAME);
-            if pool_bin_dir.exists() {
-                tokio::fs::remove_dir_all(&pool_bin_dir)
-                    .await
-                    .expect("persisted pool metadata object dir should be removable");
-                deleted_any = true;
-            }
-        }
-        // The meta-bucket layout may nest objects per pool; fall back to removing
-        // every pool.bin object directory below the temp root.
-        if !deleted_any {
-            panic!("no pool.bin found under {:?}", temp_dir.path());
-        }
-
-        let merged_newer = store
-            .reload_pool_meta()
-            .await
-            .expect("reload with missing metadata should fail closed, not error");
-        assert!(!merged_newer, "missing persisted metadata must not count as merged state");
-
-        let pool_meta = store.pool_meta.read().await;
-        let info = pool_meta.pools[0]
-            .decommission
-            .as_ref()
-            .expect("missing persisted metadata must not default local state away");
-        assert!(info.complete);
-        assert_eq!(pool_meta.pools[0].last_update, kept_time);
-
-        shutdown.cancel();
-    }
 }

crates/heal/src/heal/erasure_healer.rs

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

crates/heal/src/heal/manager.rs

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

crates/heal/src/heal/progress.rs

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

crates/heal/src/heal/resume.rs

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

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

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

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

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

crates/heal/src/heal/storage.rs

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

crates/heal/src/heal/task.rs

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

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

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

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

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

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

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

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

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

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

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

rustfs/src/storage/rpc/node_service.rs

@@ -1987,10 +1987,8 @@ impl Node for NodeService {
                 error_info: Some("errServerNotInitialized".to_string()),
             }));
         };
-        // Recover missing workers only after the reload merged newer state; a
-        // stale or duplicate reload must not spawn workers for an older generation.
         match store.reload_pool_meta().await {
-            Ok(true) => match store.spawn_missing_local_decommission_routines().await {
+            Ok(_) => match store.spawn_missing_local_decommission_routines().await {
                 Ok(_) => Ok(Response::new(ReloadPoolMetaResponse {
                     success: true,
                     error_info: None,
@@ -2000,10 +1998,6 @@ impl Node for NodeService {
                     error_info: Some(err.to_string()),
                 })),
             },
-            Ok(false) => Ok(Response::new(ReloadPoolMetaResponse {
-                success: true,
-                error_info: None,
-            })),
             Err(err) => Ok(Response::new(ReloadPoolMetaResponse {
                 success: false,
                 error_info: Some(err.to_string()),