Merge branch 'main' into cxymds/fix-1934-orphan-discovery

SSE and compression wrap the payload so its length is unknown and
advertise HashReader::SIZE_PRESERVE_LAYER (-1). Every layer preserved
that sentinel except create_bitrot_writer, which clamped it to 0 before
calling DiskAPI::create_file. RemoteDisk forwards that size verbatim in
the put_file_stream query, so remote peers were told the body was empty.

Since the authenticated put-file trailer (#5868) the receiver used the
declared size to split body from trailer, turning the clamp into a fatal
"auth trailer has trailing data" failure for every SSE PUT on multi-node
deployments (rc.2). #6320 relaxed the receiver to only trust size > 0;
this change fixes the sender so the sentinel survives end to end and the
wire no longer conflates empty objects with unknown-length streams.

Refs #6331

CLAUDE.md

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

CONTRIBUTING.md

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

Cargo.lock

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

Cargo.toml

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

crates/data-usage/src/data_usage.rs

@@ -236,15 +236,6 @@ pub struct DataUsageInfo {
     /// without relying on synchronized clocks.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage_snapshot_authoritative_baseline: Option<DataUsageSnapshotIdentity>,
-    /// Per-set freshness for an observational aggregate.  A set entry is
-    /// never sufficient to make the aggregate authoritative; it only records
-    /// which last-known-good generation contributed to the view.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub usage_snapshot_set_states: Vec<DataUsageSnapshotSetState>,
-    /// An observational view may contain only the sets that completed this
-    /// cycle (or retained a compatible last-known-good cache).
-    #[serde(default)]
-    pub usage_snapshot_partial: bool,
     /// Deprecated kept here for backward compatibility reasons
     pub bucket_sizes: HashMap<String, u64>,
     /// Per-disk snapshot information when available
@@ -261,22 +252,6 @@ pub struct DataUsageSnapshotIdentity {
     pub scanner_epoch: Option<u64>,
 }
 
-#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
-pub struct DataUsageSnapshotSetState {
-    pub pool_index: u64,
-    pub set_index: u64,
-    #[serde(default)]
-    pub scanner_cycle: Option<u64>,
-    #[serde(default)]
-    pub scanner_epoch: Option<u64>,
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub scan_plan_digest: Option<[u8; 32]>,
-    #[serde(default)]
-    pub complete: bool,
-    #[serde(default)]
-    pub tombstone: bool,
-}
-
 impl DataUsageInfo {
     pub fn snapshot_identity(&self) -> DataUsageSnapshotIdentity {
         DataUsageSnapshotIdentity {
@@ -316,7 +291,7 @@ pub fn data_usage_snapshot_is_newer(candidate: &DataUsageInfo, baseline: &DataUs
 /// rollback delete/recreate fences the previous bucket incarnation too.
 pub fn observed_data_usage_is_newer(observed: &DataUsageInfo, authoritative: &DataUsageInfo) -> bool {
     observed.usage_snapshot_converged == Some(false)
-        && (observed.is_complete_bucket_usage_snapshot() || observed.is_valid_partial_snapshot())
+        && observed.is_complete_bucket_usage_snapshot()
         && observed.usage_snapshot_authoritative_baseline.as_ref() == Some(&authoritative.snapshot_identity())
         && data_usage_snapshot_is_newer(observed, authoritative)
 }
@@ -1461,39 +1436,6 @@ impl DataUsageInfo {
             && u64::try_from(self.buckets_usage.len()).ok() == Some(self.buckets_count)
     }
 
-    /// Validate provenance before an observational view can be selected for
-    /// admin display. Partial data is accepted only with unique set states,
-    /// a plan digest for every state, and at least one usable generation.
-    pub fn is_valid_partial_snapshot(&self) -> bool {
-        if !self.usage_snapshot_partial
-            || self.usage_snapshot_converged != Some(false)
-            || self.last_update.is_none()
-            || self.scanner_cycle.is_none()
-            || self.scanner_epoch.is_none()
-            || self.usage_snapshot_set_states.is_empty()
-            || u64::try_from(self.buckets_usage.len()).ok() != Some(self.buckets_count)
-        {
-            return false;
-        }
-
-        let mut previous = None;
-        let mut plan_digest = None;
-        let mut has_source = false;
-        for state in &self.usage_snapshot_set_states {
-            if state.scan_plan_digest.is_none()
-                || plan_digest.is_some_and(|digest| Some(digest) != state.scan_plan_digest)
-                || state.scanner_cycle.is_some() != state.scanner_epoch.is_some()
-                || previous.is_some_and(|(pool, set)| (pool, set) >= (state.pool_index, state.set_index))
-            {
-                return false;
-            }
-            previous = Some((state.pool_index, state.set_index));
-            plan_digest = state.scan_plan_digest;
-            has_source |= state.scanner_cycle.is_some() && !state.tombstone;
-        }
-        has_source
-    }
-
     /// Add object metadata to data usage statistics
     pub fn add_object(&mut self, object_path: &str, meta_object: &rustfs_filemeta::MetaObject) {
         // This method is kept for backward compatibility
@@ -2321,55 +2263,6 @@ mod tests {
         assert!(!observed_data_usage_is_newer(&candidate(2, 9, Some(false), true), &authoritative));
         assert!(!observed_data_usage_is_newer(&candidate(2, 11, Some(true), true), &authoritative));
         assert!(!observed_data_usage_is_newer(&candidate(2, 11, Some(false), false), &authoritative));
-
-        let mut partial = candidate(2, 11, Some(false), false);
-        partial.usage_snapshot_partial = true;
-        partial.usage_snapshot_set_states = vec![DataUsageSnapshotSetState {
-            pool_index: 0,
-            set_index: 0,
-            scanner_cycle: Some(10),
-            scanner_epoch: Some(2),
-            scan_plan_digest: Some([1; 32]),
-            complete: false,
-            tombstone: false,
-        }];
-        assert!(observed_data_usage_is_newer(&partial, &authoritative));
-    }
-
-    #[test]
-    fn mixed_topology_snapshot_is_rejected() {
-        let mut partial = DataUsageInfo {
-            last_update: Some(SystemTime::UNIX_EPOCH + Duration::from_secs(2)),
-            scanner_cycle: Some(11),
-            scanner_epoch: Some(2),
-            buckets_count: 0,
-            usage_snapshot_converged: Some(false),
-            usage_snapshot_partial: true,
-            usage_snapshot_set_states: vec![
-                DataUsageSnapshotSetState {
-                    pool_index: 0,
-                    set_index: 0,
-                    scanner_cycle: Some(11),
-                    scanner_epoch: Some(2),
-                    scan_plan_digest: Some([1; 32]),
-                    complete: true,
-                    tombstone: false,
-                },
-                DataUsageSnapshotSetState {
-                    pool_index: 1,
-                    set_index: 0,
-                    scanner_cycle: Some(10),
-                    scanner_epoch: Some(2),
-                    scan_plan_digest: Some([2; 32]),
-                    complete: false,
-                    tombstone: false,
-                },
-            ],
-            ..Default::default()
-        };
-        assert!(!partial.is_valid_partial_snapshot());
-        partial.usage_snapshot_set_states[1].scan_plan_digest = Some([1; 32]);
-        assert!(partial.is_valid_partial_snapshot());
     }
 
     #[test]

crates/ecstore/src/data_usage/mod.rs

@@ -73,16 +73,6 @@ struct CachedBucketUsage {
     // mutation. A strictly later generation is required before the mutation
     // evidence can be discarded.
     pending_scanner_position: Option<(u64, u64)>,
-    // Deletes are visible to admin immediately, but quota admission keeps
-    // them pending until a complete scanner generation reconciles the set.
-    // This marker intentionally remains process-local: the delete request
-    // updates this overlay before the scanner writes a durable snapshot. If
-    // the process restarts first, loading the persisted complete snapshot
-    // restores the pre-reconciliation (larger) baseline, which is
-    // conservative for quota admission. A persisted post-delete snapshot is
-    // necessarily a complete scanner reconciliation and therefore creates a
-    // fresh cache entry with no pending hold.
-    pending_negative_delta: u64,
 }
 
 type UsageMemoryCache = Arc<RwLock<HashMap<String, CachedBucketUsage>>>;
@@ -958,12 +948,7 @@ async fn load_observed_data_usage_snapshot(store: Arc<ECStore>) -> Option<DataUs
     };
 
     match parse_usage_snapshot(&data) {
-        Ok(info)
-            if info.usage_snapshot_converged == Some(false)
-                && (info.is_complete_bucket_usage_snapshot() || info.is_valid_partial_snapshot()) =>
-        {
-            Some(info)
-        }
+        Ok(info) if info.usage_snapshot_converged == Some(false) && info.is_complete_bucket_usage_snapshot() => Some(info),
         Ok(_) => {
             error!(
                 event = "data_usage_snapshot_load_failed",
@@ -1008,7 +993,7 @@ async fn load_admin_data_usage_from_backend(store: Arc<ECStore>) -> Result<DataU
 }
 
 fn discard_incomplete_bucket_usage(data_usage_info: &mut DataUsageInfo) {
-    if !data_usage_info.is_complete_bucket_usage_snapshot() && !data_usage_info.usage_snapshot_partial {
+    if !data_usage_info.is_complete_bucket_usage_snapshot() {
         data_usage_info.usage_snapshot_complete = false;
         data_usage_info.buckets_usage.clear();
         data_usage_info.bucket_sizes.clear();
@@ -1658,7 +1643,6 @@ fn cached_bucket_usage_from_backend(usage: BucketUsageInfo, updated_at: SystemTi
         dirty: false,
         stale_snapshot_pending: false,
         pending_scanner_position: None,
-        pending_negative_delta: 0,
     }
 }
 
@@ -1672,7 +1656,6 @@ fn cached_bucket_usage_now(usage: BucketUsageInfo) -> CachedBucketUsage {
         dirty: false,
         stale_snapshot_pending: false,
         pending_scanner_position: None,
-        pending_negative_delta: 0,
     }
 }
 
@@ -1825,7 +1808,6 @@ pub async fn record_bucket_object_delete_memory(bucket: &str, deleted_size: u64,
         .or_insert_with(|| cached_bucket_usage_now(BucketUsageInfo::default()));
 
     entry.usage.size = entry.usage.size.saturating_sub(deleted_size);
-    entry.pending_negative_delta = entry.pending_negative_delta.saturating_add(deleted_size);
     if removed_current_object {
         entry.usage.objects_count = entry.usage.objects_count.saturating_sub(1);
         entry.usage.versions_count = entry.usage.versions_count.saturating_sub(1);
@@ -1881,7 +1863,7 @@ pub async fn get_bucket_usage_memory(bucket: &str) -> Option<u64> {
     cache
         .get(bucket)
         .filter(|cached| cached.authoritative)
-        .map(|cached| cached.usage.size.saturating_add(cached.pending_negative_delta))
+        .map(|cached| cached.usage.size)
 }
 
 async fn update_usage_cache_if_needed() {
@@ -2961,45 +2943,6 @@ mod tests {
         assert_eq!(selected.usage_snapshot_converged, Some(true));
     }
 
-    #[test]
-    fn persisted_authoritative_stalls_but_memory_overlay_remains_visible() {
-        let authoritative = DataUsageInfo {
-            last_update: Some(SystemTime::UNIX_EPOCH),
-            scanner_epoch: Some(4),
-            scanner_cycle: Some(10),
-            usage_snapshot_complete: true,
-            ..Default::default()
-        };
-        let mut partial = authoritative.clone();
-        partial.last_update = Some(SystemTime::UNIX_EPOCH + Duration::from_secs(1));
-        partial.scanner_cycle = Some(11);
-        partial.usage_snapshot_complete = false;
-        partial.usage_snapshot_partial = true;
-        partial.usage_snapshot_converged = Some(false);
-        partial.usage_snapshot_authoritative_baseline = Some(authoritative.snapshot_identity());
-        partial.usage_snapshot_set_states = vec![rustfs_data_usage::DataUsageSnapshotSetState {
-            pool_index: 0,
-            set_index: 0,
-            scanner_cycle: Some(10),
-            scanner_epoch: Some(4),
-            scan_plan_digest: Some([1; 32]),
-            complete: false,
-            tombstone: false,
-        }];
-        partial.buckets_usage.insert(
-            "bucket".to_string(),
-            BucketUsageInfo {
-                size: 100,
-                ..Default::default()
-            },
-        );
-        partial.buckets_count = 1;
-
-        let (selected, _) = select_admin_data_usage_snapshot(authoritative, true, Some(partial));
-        assert!(selected.usage_snapshot_partial);
-        assert_eq!(selected.buckets_usage.get("bucket").map(|usage| usage.size), Some(100));
-    }
-
     #[tokio::test]
     async fn authoritative_save_cleanup_removes_observed_snapshot_best_effort() {
         let store = UsageCasStore::default();
@@ -4722,55 +4665,6 @@ mod tests {
         );
     }
 
-    #[tokio::test]
-    #[serial]
-    async fn partial_usage_is_observational_not_authoritative_for_quota() {
-        clear_usage_memory_cache_for_test().await;
-
-        let mut partial = data_usage_info_for_test("bucket-a", 10, 100, SystemTime::now());
-        partial.usage_snapshot_complete = false;
-        partial.usage_snapshot_partial = true;
-        replace_bucket_usage_memory_from_info(&partial).await;
-
-        assert_eq!(get_bucket_usage_memory("bucket-a").await, None);
-    }
-
-    #[tokio::test]
-    #[serial]
-    async fn stale_quota_uses_complete_baseline_plus_positive_deltas() {
-        clear_usage_memory_cache_for_test().await;
-
-        let baseline = data_usage_info_for_test("bucket-a", 1, 100, SystemTime::now());
-        replace_bucket_usage_memory_from_info(&baseline).await;
-        record_bucket_object_write_memory("bucket-a", None, 25).await;
-
-        assert_eq!(get_bucket_usage_memory("bucket-a").await, Some(125));
-    }
-
-    #[tokio::test]
-    #[serial]
-    async fn negative_delta_waits_for_set_reconciliation() {
-        clear_usage_memory_cache_for_test().await;
-
-        let baseline = data_usage_info_for_test("bucket-a", 1, 100, SystemTime::UNIX_EPOCH + Duration::from_secs(100));
-        replace_bucket_usage_memory_from_info(&baseline).await;
-        record_bucket_object_delete_memory("bucket-a", 25, true).await;
-
-        assert_eq!(get_bucket_usage_memory("bucket-a").await, Some(100));
-
-        // Simulate a process restart: the request-path overlay is gone, but
-        // the persisted authoritative snapshot is still the pre-reconciliation
-        // baseline. Quota must remain conservative until a complete scanner
-        // result proves the delete.
-        clear_usage_memory_cache_for_test().await;
-        replace_bucket_usage_memory_from_info(&baseline).await;
-        assert_eq!(get_bucket_usage_memory("bucket-a").await, Some(100));
-
-        let reconciled = data_usage_info_for_test("bucket-a", 0, 75, SystemTime::UNIX_EPOCH + Duration::from_secs(101));
-        replace_bucket_usage_memory_from_info(&reconciled).await;
-        assert_eq!(get_bucket_usage_memory("bucket-a").await, Some(75));
-    }
-
     #[tokio::test]
     #[serial]
     async fn memory_overlay_counts_versioned_overwrite_as_new_version() {

crates/ecstore/src/io_support/bitrot.rs

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

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

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

crates/ecstore/src/store/object.rs

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

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

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

crates/filemeta/src/metacache.rs

@@ -21,6 +21,7 @@ use arc_swap::ArcSwapOption;
 use rmp::Marker;
 use serde::{Deserialize, Serialize};
 use std::cmp::Ordering;
+use std::collections::{HashMap, HashSet};
 use std::str::from_utf8;
 use std::{
     fmt::Debug,
@@ -37,8 +38,13 @@ use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
 use tokio::spawn;
 use tokio::sync::Mutex;
 use tracing::{debug, warn};
+use uuid::Uuid;
 
 const SLASH_SEPARATOR: &str = "/";
+pub const MAX_META_CACHE_HEAL_CANDIDATES: usize = 1024;
+/// Keep truncation continuations bounded while still giving the scanner a
+/// safe object-level retry for versions that did not fit in the candidate set.
+pub const MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS: usize = 64;
 
 #[derive(Clone, Debug, Default)]
 pub struct MetadataResolutionParams {
@@ -66,6 +72,50 @@ pub struct MetaCacheEntry {
     pub reusable: bool,
 }
 
+#[derive(Clone, Debug, PartialEq, Eq, Hash)]
+pub enum MetaCacheHealCandidateKind {
+    Object,
+    DeleteMarker,
+    UnversionedObject,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Hash)]
+pub struct MetaCacheHealCandidate {
+    pub object: String,
+    pub version_id: Option<Uuid>,
+    pub kind: MetaCacheHealCandidateKind,
+    /// Number of raw disk entries that carried this validated version.
+    pub replica_count: usize,
+}
+
+impl MetaCacheHealCandidate {
+    pub fn validated_version(&self) -> Option<Uuid> {
+        match self.kind {
+            MetaCacheHealCandidateKind::Object | MetaCacheHealCandidateKind::DeleteMarker => self.version_id,
+            MetaCacheHealCandidateKind::UnversionedObject => None,
+        }
+    }
+
+    pub fn is_unversioned(&self) -> bool {
+        self.kind == MetaCacheHealCandidateKind::UnversionedObject
+    }
+}
+
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub struct MetaCacheHealDiscovery {
+    pub candidates: Vec<MetaCacheHealCandidate>,
+    pub unverified_count: usize,
+    pub truncated: bool,
+    /// Object names whose validated version set exceeded the candidate cap.
+    /// The scanner retries these names without a version and with destructive
+    /// healing disabled; this is an explicit bounded continuation, not a
+    /// version claim.
+    pub truncated_objects: Vec<String>,
+    /// Validated candidates beyond the main cap, retained with exact version
+    /// identities so callers never fall back to a latest-version request.
+    pub truncated_candidates: Vec<MetaCacheHealCandidate>,
+}
+
 impl MetaCacheEntry {
     pub fn marshal_msg(&self) -> Result<Vec<u8>> {
         let mut wr = Vec::new();
@@ -370,6 +420,185 @@ impl MetaCacheEntries {
         })
     }
 
+    /// Discover validated object/delete-marker versions and safe unversioned
+    /// inspection candidates in the raw entries without applying read quorum.
+    /// This is intentionally separate from [`Self::resolve`]: a sub-quorum
+    /// version is a valid heal target even though it must not participate in
+    /// normal reads or writes.
+    ///
+    /// The validated list is bounded and deduplicated by object, version id,
+    /// and metadata kind; each candidate retains the number of raw disk
+    /// entries that carried it so callers can classify sub-quorum versions.
+    /// Entries whose xl.meta cannot be decoded are counted separately for
+    /// discovery accounting; they never become versionless destructive heal
+    /// requests and do not consume the validated quota. An
+    /// [`MetaCacheHealCandidateKind::UnversionedObject`] is always consumed by
+    /// a non-destructive scanner request.
+    pub fn discover_heal_candidates(&self, bucket: &str, max_candidates: usize) -> MetaCacheHealDiscovery {
+        let limit = max_candidates.min(MAX_META_CACHE_HEAL_CANDIDATES);
+        if limit == 0 || bucket.is_empty() {
+            return MetaCacheHealDiscovery::default();
+        }
+
+        let mut discovery = MetaCacheHealDiscovery {
+            candidates: Vec::<MetaCacheHealCandidate>::with_capacity(limit.min(self.0.len())),
+            unverified_count: 0,
+            truncated: false,
+            truncated_objects: Vec::with_capacity(MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS.min(limit)),
+            truncated_candidates: Vec::new(),
+        };
+        let mut seen: HashMap<(String, Option<Uuid>, MetaCacheHealCandidateKind), usize> =
+            HashMap::with_capacity(limit.min(self.0.len()));
+
+        for entry in self.0.iter().flatten() {
+            if !valid_heal_candidate_name(bucket, entry) {
+                continue;
+            }
+
+            let meta = match FileMeta::load(&entry.metadata) {
+                Ok(meta) => meta,
+                Err(_) => {
+                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                    continue;
+                }
+            };
+            let mut entry_seen = HashSet::new();
+
+            for shallow in meta.versions {
+                let version = match shallow.parse_version_meta() {
+                    Ok(version) if version.valid() => version,
+                    Ok(_) | Err(_) => {
+                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                        continue;
+                    }
+                };
+                if version.free_version() {
+                    continue;
+                }
+
+                let payload_header = version.header();
+                if normalize_version_id(shallow.header.version_id) != normalize_version_id(payload_header.version_id)
+                    || shallow.header.version_type != payload_header.version_type
+                {
+                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                    continue;
+                }
+
+                let (kind, version_id) = match version.version_type {
+                    VersionType::Object
+                        if version.object.is_some() && version.delete_marker.is_none() && version.legacy_object.is_none() =>
+                    {
+                        match version.object.as_ref().and_then(|object| object.version_id) {
+                            Some(id) if !id.is_nil() => (MetaCacheHealCandidateKind::Object, Some(id)),
+                            Some(_) | None => (MetaCacheHealCandidateKind::UnversionedObject, None),
+                        }
+                    }
+                    VersionType::Delete
+                        if version.delete_marker.is_some() && version.object.is_none() && version.legacy_object.is_none() =>
+                    {
+                        let Some(id) = version.delete_marker.as_ref().and_then(|marker| marker.version_id) else {
+                            discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                            continue;
+                        };
+                        if id.is_nil() {
+                            discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                            continue;
+                        }
+                        (MetaCacheHealCandidateKind::DeleteMarker, Some(id))
+                    }
+                    VersionType::Legacy
+                        if version.legacy_object.is_some() && version.object.is_none() && version.delete_marker.is_none() =>
+                    {
+                        let Some(legacy) = version.legacy_object.as_ref() else {
+                            continue;
+                        };
+                        if legacy.version_id.is_empty() {
+                            (MetaCacheHealCandidateKind::UnversionedObject, None)
+                        } else {
+                            let Ok(id) = Uuid::parse_str(&legacy.version_id) else {
+                                discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                                continue;
+                            };
+                            if id.is_nil() {
+                                discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                                continue;
+                            }
+                            (MetaCacheHealCandidateKind::Object, Some(id))
+                        }
+                    }
+                    _ => {
+                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                        continue;
+                    }
+                };
+
+                if normalize_version_id(payload_header.version_id) != version_id {
+                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                    continue;
+                }
+
+                // `all_parts=true` is the trust-boundary check for versioned
+                // candidates. A null/legacy object may still need the old
+                // non-destructive inspection fallback when its part arrays
+                // are parseable but incomplete; never use that fallback for
+                // a candidate carrying a real version id.
+                let file_info = match version.clone().into_fileinfo(bucket, &entry.name, true) {
+                    Ok(file_info) => file_info,
+                    Err(_) if version_id.is_none() && matches!(kind, MetaCacheHealCandidateKind::UnversionedObject) => {
+                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                        match version.into_fileinfo(bucket, &entry.name, false) {
+                            Ok(file_info) => file_info,
+                            Err(_) => continue,
+                        }
+                    }
+                    Err(_) => {
+                        discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                        continue;
+                    }
+                };
+                if file_info.volume != bucket || file_info.name != entry.name {
+                    discovery.unverified_count = discovery.unverified_count.saturating_add(1);
+                    continue;
+                }
+
+                let candidate = MetaCacheHealCandidate {
+                    object: entry.name.clone(),
+                    version_id,
+                    kind,
+                    replica_count: 1,
+                };
+                let key = (candidate.object.clone(), candidate.version_id, candidate.kind.clone());
+                if entry_seen.contains(&key) {
+                    continue;
+                }
+                if let Some(index) = seen.get(&key).copied() {
+                    entry_seen.insert(key);
+                    discovery.candidates[index].replica_count = discovery.candidates[index].replica_count.saturating_add(1);
+                } else if discovery.candidates.len() >= limit {
+                    // Keep the validated candidate list bounded, but retain a
+                    // bounded object-level continuation so the scanner cannot
+                    // silently lose every version of a busy object.
+                    discovery.truncated = true;
+                    if discovery.truncated_objects.len() < MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
+                        && !discovery.truncated_objects.iter().any(|object| object == &candidate.object)
+                    {
+                        discovery.truncated_objects.push(candidate.object.clone());
+                    }
+                    if discovery.truncated_objects.iter().any(|object| object == &candidate.object) {
+                        discovery.truncated_candidates.push(candidate);
+                    }
+                    continue;
+                } else {
+                    entry_seen.insert(key.clone());
+                    seen.insert(key, discovery.candidates.len());
+                    discovery.candidates.push(candidate);
+                }
+            }
+        }
+
+        discovery
+    }
+
     fn resolve_inner(&self, mut params: MetadataResolutionParams, enforce_write_quorum: bool) -> Option<MetaCacheEntry> {
         if self.0.is_empty() {
             debug!(
@@ -546,6 +775,33 @@ impl MetaCacheEntries {
     }
 }
 
+fn valid_heal_candidate_name(bucket: &str, entry: &MetaCacheEntry) -> bool {
+    if bucket.is_empty()
+        || entry.name.is_empty()
+        || entry.is_dir()
+        || (cfg!(windows) && entry.name.contains('\\'))
+        || entry.name.chars().any(char::is_control)
+    {
+        return false;
+    }
+
+    // Validate raw key components without normalizing them. The scanner maps
+    // accepted keys to filesystem paths later, so dot components and empty
+    // internal components must be rejected before that boundary. A final
+    // empty component is retained for valid keys ending in '/'.
+    let mut components = entry.name.split('/').peekable();
+    while let Some(component) = components.next() {
+        if component == "." || component == ".." || (component.is_empty() && components.peek().is_some()) {
+            return false;
+        }
+    }
+    true
+}
+
+fn normalize_version_id(version_id: Option<Uuid>) -> Option<Uuid> {
+    version_id.filter(|id| !id.is_nil())
+}
+
 #[derive(Debug, Default)]
 pub struct MetaCacheEntriesSortedResult {
     pub entries: Option<MetaCacheEntriesSorted>,
@@ -991,7 +1247,7 @@ impl<T: Clone + Debug + Send + Sync + 'static> Cache<T> {
 mod tests {
     use super::*;
     use crate::test_data::create_real_xlmeta;
-    use crate::{FileMetaVersion, MetaDeleteMarker, TRANSITION_COMPLETE};
+    use crate::{FileMetaVersion, MetaDeleteMarker, MetaObjectV1, MetaObjectV1Erasure, MetaObjectV1Stat, TRANSITION_COMPLETE};
     use std::collections::HashMap;
     use std::io::Cursor;
     use std::sync::{
@@ -1592,6 +1848,381 @@ mod tests {
         );
     }
 
+    #[test]
+    fn discover_heal_candidates_keeps_sub_quorum_versions_and_deduplicates() {
+        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+        let entries = MetaCacheEntries(vec![
+            Some(metacache_entry_single_version(1, now, "one")),
+            Some(metacache_entry_single_version(2, now, "two")),
+            Some(metacache_entry_single_version(2, now, "two")),
+            Some(metacache_entry_single_version(3, now, "three")),
+        ]);
+
+        let discovery = entries.discover_heal_candidates("bucket", 16);
+        let ids: std::collections::HashSet<Uuid> = discovery
+            .candidates
+            .iter()
+            .filter_map(|candidate| candidate.version_id)
+            .collect();
+        assert_eq!(
+            ids,
+            [Uuid::from_u128(1), Uuid::from_u128(2), Uuid::from_u128(3)]
+                .into_iter()
+                .collect()
+        );
+        assert_eq!(discovery.candidates.len(), 3, "duplicate tied versions must be emitted once");
+        assert_eq!(
+            discovery
+                .candidates
+                .iter()
+                .find(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
+                .expect("duplicate version should be discovered")
+                .replica_count,
+            2
+        );
+    }
+
+    #[test]
+    fn discover_heal_candidates_does_not_count_duplicate_versions_within_one_entry() {
+        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+        let mut meta = FileMeta::load(&metacache_entry_single_version(1, now, "duplicate").metadata)
+            .expect("duplicate fixture should decode");
+        meta.versions.push(meta.versions[0].clone());
+        let entry = MetaCacheEntry {
+            name: "object".to_string(),
+            metadata: meta.marshal_msg().expect("duplicate metadata should marshal"),
+            cached: Some(meta),
+            reusable: false,
+        };
+
+        let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 16);
+        let candidate = discovery
+            .candidates
+            .iter()
+            .find(|candidate| candidate.version_id == Some(Uuid::from_u128(1)))
+            .expect("duplicate fixture should be discovered");
+        assert_eq!(candidate.replica_count, 1);
+    }
+
+    #[test]
+    fn discover_heal_candidates_covers_divergent_quorum_boundaries_n2_n4_n6() {
+        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+
+        for (disk_count, quorum) in [(2usize, 1usize), (4, 2), (6, 3)] {
+            let target_id = Uuid::from_u128(0x1000 + disk_count as u128);
+            for target_replicas in [quorum.saturating_sub(1), quorum, quorum + 1] {
+                let entries = (0..disk_count)
+                    .map(|disk| {
+                        let version_id = if disk < target_replicas {
+                            target_id
+                        } else {
+                            Uuid::from_u128(0x2000 + disk as u128)
+                        };
+                        Some(metacache_entry_single_version(version_id.as_u128(), now, "divergent"))
+                    })
+                    .collect();
+                let discovery = MetaCacheEntries(entries).discover_heal_candidates("bucket", 32);
+                let target = discovery
+                    .candidates
+                    .iter()
+                    .find(|candidate| candidate.version_id == Some(target_id));
+                assert_eq!(target.is_some(), target_replicas > 0, "N={disk_count}, replicas={target_replicas}");
+                if let Some(target) = target {
+                    assert_eq!(target.replica_count, target_replicas);
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn discover_heal_candidates_separates_delete_markers_and_preserves_unversioned_objects() {
+        let mut marker_meta = FileMeta::new();
+        marker_meta
+            .add_version(FileInfo {
+                volume: "bucket".to_string(),
+                name: "object".to_string(),
+                version_id: Some(Uuid::from_u128(99)),
+                deleted: true,
+                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
+                ..Default::default()
+            })
+            .expect("delete marker should be added");
+        let marker = MetaCacheEntry {
+            name: "object".to_string(),
+            metadata: marker_meta.marshal_msg().expect("delete marker metadata should marshal"),
+            cached: Some(marker_meta),
+            reusable: false,
+        };
+
+        let unversioned_entry = metacache_entry_with_mod_time(
+            OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp"),
+            "unversioned",
+        );
+        let discovery = MetaCacheEntries(vec![Some(marker), Some(unversioned_entry)]).discover_heal_candidates("bucket", 16);
+        assert!(discovery.candidates.iter().any(|candidate| {
+            candidate.kind == MetaCacheHealCandidateKind::DeleteMarker && candidate.version_id == Some(Uuid::from_u128(99))
+        }));
+        assert!(discovery.candidates.iter().any(|candidate| {
+            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
+        }));
+    }
+
+    #[test]
+    fn discover_heal_candidates_rejects_delete_markers_without_ids() {
+        let mut marker_meta = FileMeta::new();
+        marker_meta
+            .add_version(FileInfo {
+                volume: "bucket".to_string(),
+                name: "object".to_string(),
+                deleted: true,
+                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
+                ..Default::default()
+            })
+            .expect("nil delete marker should be added");
+        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
+            name: "object".to_string(),
+            metadata: marker_meta.marshal_msg().expect("nil marker metadata should marshal"),
+            cached: Some(marker_meta),
+            reusable: false,
+        })])
+        .discover_heal_candidates("bucket", 16);
+
+        assert!(
+            !discovery
+                .candidates
+                .iter()
+                .any(|candidate| candidate.kind == MetaCacheHealCandidateKind::DeleteMarker)
+        );
+        assert!(discovery.unverified_count >= 1);
+    }
+
+    #[test]
+    fn discover_heal_candidates_skips_free_versions() {
+        let object_id = Uuid::from_u128(100);
+        let free_id = Uuid::from_u128(101);
+        let mut meta = FileMeta::new();
+        meta.add_version(FileInfo {
+            volume: "bucket".to_string(),
+            name: "object".to_string(),
+            version_id: Some(object_id),
+            transition_status: TRANSITION_COMPLETE.to_string(),
+            transitioned_objname: "remote/object".to_string(),
+            transition_version_id: Some(Uuid::from_u128(102)),
+            transition_tier: "WARM".to_string(),
+            mod_time: Some(OffsetDateTime::now_utc()),
+            ..Default::default()
+        })
+        .expect("transitioned object should be added");
+        let mut delete = FileInfo {
+            volume: "bucket".to_string(),
+            name: "object".to_string(),
+            version_id: Some(object_id),
+            mod_time: Some(OffsetDateTime::now_utc()),
+            ..Default::default()
+        };
+        delete.set_tier_free_version_id(&free_id.to_string());
+        meta.delete_version(&delete).expect("free version should be persisted");
+
+        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
+            name: "object".to_string(),
+            metadata: meta.marshal_msg().expect("free version metadata should marshal"),
+            cached: Some(meta),
+            reusable: false,
+        })])
+        .discover_heal_candidates("bucket", 16);
+        assert!(discovery.candidates.is_empty());
+    }
+
+    #[test]
+    fn discover_heal_candidates_preserves_unversioned_legacy_object() {
+        let legacy = MetaObjectV1 {
+            version: "1.0.1".to_string(),
+            format: "xl".to_string(),
+            stat: MetaObjectV1Stat {
+                size: 1,
+                mod_time: Some(OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp")),
+                name: "object".to_string(),
+                ..Default::default()
+            },
+            erasure: MetaObjectV1Erasure {
+                data_blocks: 4,
+                parity_blocks: 2,
+                index: 1,
+                distribution: vec![1, 2, 3, 4, 5, 6],
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+        let version = FileMetaVersion {
+            version_type: VersionType::Legacy,
+            legacy_object: Some(legacy),
+            ..Default::default()
+        };
+        let mut meta = FileMeta::new();
+        meta.versions
+            .push(FileMetaShallowVersion::try_from(version).expect("legacy metadata should marshal"));
+        let discovery = MetaCacheEntries(vec![Some(MetaCacheEntry {
+            name: "object".to_string(),
+            metadata: meta.marshal_msg().expect("legacy metadata should marshal"),
+            cached: Some(meta),
+            reusable: false,
+        })])
+        .discover_heal_candidates("bucket", 16);
+        assert!(discovery.candidates.iter().any(|candidate| {
+            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
+        }));
+    }
+
+    #[test]
+    fn discover_heal_candidates_rejects_nil_and_malformed_metadata_and_is_bounded() {
+        let now = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+        let mut nil = metacache_entry_single_version(1, now, "nil");
+        let mut nil_meta = FileMeta::load(&nil.metadata).expect("nil fixture should decode");
+        let mut nil_version = nil_meta.versions[0]
+            .parse_version_meta()
+            .expect("nil fixture version should decode");
+        nil_version.object.as_mut().expect("object fixture").version_id = Some(Uuid::nil());
+        nil_meta.versions[0] = FileMetaShallowVersion::try_from(nil_version).expect("nil fixture should marshal");
+        nil.metadata = nil_meta.marshal_msg().expect("nil fixture metadata should marshal");
+
+        let mut mismatched = metacache_entry_single_version(2, now, "mismatched");
+        let mut mismatched_meta = FileMeta::load(&mismatched.metadata).expect("mismatched fixture should decode");
+        mismatched_meta.versions[0].header.version_id = Some(Uuid::from_u128(200));
+        mismatched.metadata = mismatched_meta.marshal_msg().expect("mismatched metadata should marshal");
+
+        let mut short_parts = metacache_entry_single_version(3, now, "short-parts");
+        let mut short_parts_meta = FileMeta::load(&short_parts.metadata).expect("short-parts fixture should decode");
+        let mut short_parts_version = short_parts_meta.versions[0]
+            .parse_version_meta()
+            .expect("short-parts fixture version should decode");
+        let object = short_parts_version.object.as_mut().expect("object fixture");
+        object.part_numbers = vec![1];
+        object.part_actual_sizes = vec![1];
+        object.part_sizes.clear();
+        short_parts_meta.versions[0] = FileMetaShallowVersion::try_from(short_parts_version).expect("short-parts should marshal");
+        short_parts.metadata = short_parts_meta.marshal_msg().expect("short-parts metadata should marshal");
+
+        let mut short_unversioned = metacache_entry_with_mod_time(now, "short-unversioned");
+        let mut short_unversioned_meta =
+            FileMeta::load(&short_unversioned.metadata).expect("short-unversioned fixture should decode");
+        let mut short_unversioned_version = short_unversioned_meta.versions[0]
+            .parse_version_meta()
+            .expect("short-unversioned version should decode");
+        let unversioned_object = short_unversioned_version.object.as_mut().expect("unversioned object fixture");
+        unversioned_object.part_numbers = vec![1];
+        unversioned_object.part_actual_sizes = vec![1];
+        unversioned_object.part_sizes.clear();
+        short_unversioned_meta.versions[0] =
+            FileMetaShallowVersion::try_from(short_unversioned_version).expect("short-unversioned should marshal");
+        short_unversioned.metadata = short_unversioned_meta
+            .marshal_msg()
+            .expect("short-unversioned metadata should marshal");
+
+        let mut malformed = nil.clone();
+        malformed.name = "malformed".to_string();
+        malformed.metadata = vec![1, 2, 3];
+
+        let entries = MetaCacheEntries(
+            std::iter::once(Some(nil))
+                .chain(std::iter::once(Some(mismatched)))
+                .chain(std::iter::once(Some(short_parts)))
+                .chain(std::iter::once(Some(short_unversioned)))
+                .chain(std::iter::once(Some(malformed)))
+                .chain((0..32).map(|id| Some(metacache_entry_single_version(id + 10, now, "bounded"))))
+                .collect(),
+        );
+        let discovery = entries.discover_heal_candidates("bucket", 5);
+        assert!(discovery.candidates.len() <= 5);
+        assert!(discovery.truncated, "bounded discovery must expose dropped candidates");
+        assert!(
+            discovery
+                .truncated_candidates
+                .iter()
+                .all(|candidate| candidate.version_id.is_some()),
+            "overflow candidates must retain exact version identities"
+        );
+        assert!(
+            discovery.truncated_objects.iter().any(|object| object == "object"),
+            "bounded discovery must expose an object-level safe continuation"
+        );
+        assert!(
+            !discovery
+                .candidates
+                .iter()
+                .any(|candidate| candidate.version_id == Some(Uuid::nil()))
+        );
+        assert!(
+            !discovery
+                .candidates
+                .iter()
+                .any(|candidate| candidate.version_id == Some(Uuid::from_u128(2)))
+        );
+        assert!(
+            !discovery
+                .candidates
+                .iter()
+                .any(|candidate| candidate.version_id == Some(Uuid::from_u128(3)))
+        );
+        assert!(discovery.candidates.iter().any(|candidate| {
+            candidate.kind == MetaCacheHealCandidateKind::UnversionedObject && candidate.version_id.is_none()
+        }));
+        assert!(
+            discovery.unverified_count >= 1,
+            "malformed and rejected metadata must remain observable during discovery"
+        );
+
+        for invalid_name in [
+            "../object",
+            "./object",
+            "object/../other",
+            "object//name",
+            "object\u{0001}name",
+            "object\0name",
+        ] {
+            let mut entry = metacache_entry_single_version(400, now, invalid_name);
+            entry.name = invalid_name.to_string();
+            let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
+            assert!(
+                discovery.candidates.is_empty(),
+                "invalid key should not become a heal candidate: {invalid_name:?}"
+            );
+        }
+
+        #[cfg(windows)]
+        {
+            let mut entry = metacache_entry_single_version(400, now, "object\\name");
+            entry.name = "object\\name".to_string();
+            assert!(
+                MetaCacheEntries(vec![Some(entry)])
+                    .discover_heal_candidates("bucket", 5)
+                    .candidates
+                    .is_empty(),
+                "backslash is a path separator on Windows"
+            );
+        }
+
+        #[cfg(not(windows))]
+        {
+            let mut entry = metacache_entry_single_version(400, now, "object\\name");
+            entry.name = "object\\name".to_string();
+            assert_eq!(
+                MetaCacheEntries(vec![Some(entry)])
+                    .discover_heal_candidates("bucket", 5)
+                    .candidates
+                    .len(),
+                1,
+                "backslash is object-key data on Unix"
+            );
+        }
+
+        for valid_name in ["trailing/", "prefix/object"] {
+            let mut entry = metacache_entry_single_version(401, now, valid_name);
+            entry.name = valid_name.to_string();
+            let discovery = MetaCacheEntries(vec![Some(entry)]).discover_heal_candidates("bucket", 5);
+            assert_eq!(discovery.candidates.len(), 1, "raw S3 key should remain opaque: {valid_name:?}");
+        }
+    }
+
     #[test]
     fn resolve_rejects_partial_latest_and_returns_committed_previous_metadata() {
         let old_mod_time = OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");

crates/scanner/src/data_usage_define.rs

@@ -28,9 +28,8 @@ use rustfs_common::heal_channel::HealScanMode;
 use rustfs_config::ENV_SCANNER_CACHE_SAVE_TIMEOUT_SECS;
 pub use rustfs_data_usage::{
     AllTierStats, BucketTargetUsageInfo, BucketUsageInfo, DATA_USAGE_OBJECT_NAME, DATA_USAGE_OBSERVED_OBJECT_NAME,
-    DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, DataUsageSnapshotSetState, LEGACY_DATA_USAGE_OBJECT_NAME,
-    PrefixUsageEntry, PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeSummary, TierStats, hash_path,
-    prefix_usage_in_cache,
+    DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, LEGACY_DATA_USAGE_OBJECT_NAME, PrefixUsageEntry,
+    PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeSummary, TierStats, hash_path, prefix_usage_in_cache,
 };
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
 use tokio::time::{Duration, Instant, sleep, timeout};
@@ -345,18 +344,6 @@ pub struct DataUsageCacheInfo {
     pub scan_plan_digest: Option<DataUsageScanPlanDigest>,
     #[serde(default)]
     pub cache_key_format: u16,
-    /// Whether the entries retained while a set scan was incomplete come
-    /// from a prior complete set snapshot.  This is observational input only.
-    #[serde(default)]
-    pub lkg_snapshot_complete: bool,
-    #[serde(default)]
-    pub lkg_next_cycle: Option<u64>,
-    #[serde(default)]
-    pub lkg_last_update: Option<SystemTime>,
-    #[serde(default)]
-    pub lkg_leader_epoch: Option<u64>,
-    #[serde(default)]
-    pub lkg_scan_plan_digest: Option<DataUsageScanPlanDigest>,
 }
 
 impl Serialize for DataUsageCacheInfo {
@@ -366,7 +353,7 @@ impl Serialize for DataUsageCacheInfo {
     {
         // Keep this metadata map-encoded so older readers can ignore fields
         // appended by newer scanner versions during rolling upgrades.
-        let mut state = serializer.serialize_map(Some(21))?;
+        let mut state = serializer.serialize_map(Some(16))?;
         state.serialize_entry("name", &self.name)?;
         state.serialize_entry("next_cycle", &self.next_cycle)?;
         state.serialize_entry("leader_epoch", &self.leader_epoch)?;
@@ -383,11 +370,6 @@ impl Serialize for DataUsageCacheInfo {
         state.serialize_entry("snapshot_complete", &self.snapshot_complete)?;
         state.serialize_entry("scan_plan_digest", &self.scan_plan_digest)?;
         state.serialize_entry("cache_key_format", &self.cache_key_format)?;
-        state.serialize_entry("lkg_snapshot_complete", &self.lkg_snapshot_complete)?;
-        state.serialize_entry("lkg_next_cycle", &self.lkg_next_cycle)?;
-        state.serialize_entry("lkg_last_update", &self.lkg_last_update)?;
-        state.serialize_entry("lkg_leader_epoch", &self.lkg_leader_epoch)?;
-        state.serialize_entry("lkg_scan_plan_digest", &self.lkg_scan_plan_digest)?;
         state.end()
     }
 }

crates/scanner/src/scanner.rs

@@ -2274,8 +2274,9 @@ async fn final_data_usage_publication_defer_reason(
             }
         }
         ScannerCycleStatus::Deferred(reason) => Some(reason),
-        // Incomplete cycles may publish a non-authoritative observational
-        // snapshot when at least one set has a usable current/LKG view.
+        // Incomplete cycles do not publish a usage snapshot. Keep the
+        // decision permissive so existing partial-cycle handling remains
+        // unchanged if a future scanner path emits a bookkeeping update.
         ScannerCycleStatus::Incomplete => None,
     }
 }

crates/scanner/src/scanner/usage_store.rs

@@ -198,7 +198,7 @@ where
             data_usage_info.usage_snapshot_authoritative_baseline = Some(authoritative.snapshot_identity());
         }
 
-        if !data_usage_info.is_complete_bucket_usage_snapshot() && !data_usage_info.usage_snapshot_partial {
+        if !data_usage_info.is_complete_bucket_usage_snapshot() {
             error!(
                 target: "rustfs::scanner",
                 event = EVENT_SCANNER_PERSIST_STATE,

crates/scanner/src/scanner_folder.rs

@@ -43,7 +43,10 @@ use rustfs_common::metrics::{
     UpdateCurrentPathFn, current_path_updater, global_metrics,
 };
 use rustfs_common::trace_bus::{TraceEvent, TraceFunc, TraceKind, trace_emit, trace_subscriber_count};
-use rustfs_filemeta::{MetaCacheEntries, MetaCacheEntry, MetadataResolutionParams};
+use rustfs_filemeta::{
+    MAX_META_CACHE_HEAL_CANDIDATES, MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS, MetaCacheEntries, MetaCacheEntry,
+    MetaCacheHealCandidateKind,
+};
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
 use s3s::dto::{BucketLifecycleConfiguration, ObjectLockConfiguration, VersioningConfiguration};
 use time::OffsetDateTime;
@@ -96,6 +99,11 @@ const METRIC_SCANNER_EXCESS_OBJECT_VERSION_SIZE_TOTAL: &str = "rustfs_scanner_ex
 const METRIC_SCANNER_EXCESS_FOLDERS_TOTAL: &str = "rustfs_scanner_excess_folders_total";
 const METRIC_SCANNER_PENDING_HEAL_PRUNE_TOTAL: &str = "rustfs_scanner_pending_heal_prune_total";
 const METRIC_SCANNER_PENDING_HEAL_MALFORMED_TOTAL: &str = "rustfs_scanner_pending_heal_malformed_total";
+const METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL: &str = "rustfs_scanner_heal_discovery_candidates_total";
+const METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL: &str = "rustfs_scanner_heal_discovery_sub_quorum_total";
+const METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL: &str = "rustfs_scanner_heal_discovery_unverified_total";
+const METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL: &str = "rustfs_scanner_heal_discovery_queued_total";
+const METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL: &str = "rustfs_scanner_heal_discovery_truncated_total";
 const MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET: usize = 128;
 
 // --- scanner excess alerts as S3 notification events (rustfs/backlog#1868) --
@@ -883,7 +891,7 @@ impl FolderScanner {
         object: Option<String>,
         version_id: Option<String>,
         request: HealChannelRequest,
-    ) -> Result<(), ScannerError> {
+    ) -> Result<HealAdmissionResult, ScannerError> {
         let candidate_type = pending_scanner_heal_candidate_type(kind);
         let priority = request.priority;
         let scan_mode = request.scan_mode.unwrap_or(self.scan_mode);
@@ -911,7 +919,7 @@ impl FolderScanner {
                     error = %err,
                     "Scanner deferred heal request after channel error"
                 );
-                return Ok(());
+                return Ok(HealAdmissionResult::Full);
             }
         };
         self.update_pending_scanner_heal_after_admission(
@@ -923,7 +931,7 @@ impl FolderScanner {
             result,
         );
         if result.is_admitted() {
-            return Ok(());
+            return Ok(result);
         }
 
         record_high_priority_heal_escalation(candidate_type, priority, result);
@@ -944,7 +952,7 @@ impl FolderScanner {
             state = "high_priority_not_admitted",
             "Scanner high-priority heal admission failed"
         );
-        Ok(())
+        Ok(result)
     }
 
     pub fn set_heal_object_select(&mut self, prob: u32) {
@@ -1736,14 +1744,7 @@ impl FolderScanner {
                 break;
             }
 
-            let mut resolver = MetadataResolutionParams {
-                dir_quorum: self.disks_quorum,
-                obj_quorum: self.disks_quorum,
-                bucket: "".to_string(),
-                strict: false,
-                ..Default::default()
-            };
-
+            let mut previous_bucket = String::new();
             for name in abandoned_children {
                 if !self.should_heal().await {
                     break;
@@ -1751,7 +1752,7 @@ impl FolderScanner {
 
                 let (bucket, prefix) = path2_bucket_object(name.as_str());
 
-                if bucket != resolver.bucket {
+                if bucket != previous_bucket {
                     self.send_required_scanner_heal_request(
                         PendingScannerHealKind::Bucket,
                         bucket.clone(),
@@ -1760,10 +1761,9 @@ impl FolderScanner {
                         build_bucket_heal_request(bucket.clone(), HealChannelPriority::High),
                     )
                     .await?;
+                    previous_bucket = bucket.clone();
                 }
 
-                resolver.bucket = bucket.clone();
-
                 let child_ctx = ctx.child_token();
 
                 let (agreed_tx, mut agreed_rx) = mpsc::channel::<String>(1);
@@ -1880,6 +1880,8 @@ impl FolderScanner {
                 let mut agreed_closed = false;
                 let mut partial_closed = false;
                 let mut finished_closed = false;
+                let mut seen_heal_candidates: HashSet<(String, Option<String>, MetaCacheHealCandidateKind)> = HashSet::new();
+                let mut seen_truncated_objects: HashSet<String> = HashSet::new();
 
                 loop {
                     if agreed_closed && partial_closed && finished_closed {
@@ -1904,65 +1906,112 @@ impl FolderScanner {
                                 break;
                             }
 
-                            let Some(entry) = resolve_object_heal_entry(&entries, resolver.clone()) else {
-                                continue;
-                            };
-
-                            (self.update_current_path)(&entry.name).await;
-
-                            if entry.is_dir() {
-                                continue;
+                            let discovery = entries.discover_heal_candidates(&bucket, MAX_META_CACHE_HEAL_CANDIDATES);
+                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_CANDIDATES_TOTAL)
+                                .increment(u64::try_from(discovery.candidates.len()).unwrap_or(u64::MAX));
+                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_SUB_QUORUM_TOTAL).increment(
+                                u64::try_from(
+                                    discovery
+                                        .candidates
+                                        .iter()
+                                        .filter(|candidate| candidate.replica_count < disks_quorum)
+                                        .count(),
+                                )
+                                .unwrap_or(u64::MAX),
+                            );
+                            counter!(METRIC_SCANNER_HEAL_DISCOVERY_UNVERIFIED_TOTAL).increment(
+                                u64::try_from(discovery.unverified_count).unwrap_or(u64::MAX),
+                            );
+                            if discovery.truncated {
+                                counter!(METRIC_SCANNER_HEAL_DISCOVERY_TRUNCATED_TOTAL).increment(1);
                             }
 
-                            let fivs = match entry.file_info_versions(&bucket) {
-                                Ok(fivs) => fivs,
-                                Err(e) => {
-                                    error!(
-                                        target: "rustfs::scanner::folder",
-                                        event = EVENT_SCANNER_FOLDER_STATE,
-                                        component = LOG_COMPONENT_SCANNER,
-                                        subsystem = LOG_SUBSYSTEM_FOLDER,
-                                        bucket = %bucket,
-                                        entry = %entry.name,
-                                        state = "file_info_versions_failed",
-                                        error = %e,
-                                        "Scanner list_path_raw failed to resolve file versions"
-                                    );
-                                    self.send_required_scanner_heal_request(
-                                        PendingScannerHealKind::Object,
-                                        bucket.clone(),
-                                        Some(entry.name.clone()),
-                                        None,
-                                        build_object_heal_request(
-                                            bucket.clone(),
-                                            entry.name.clone(),
-                                            None,
-                                            self.scan_mode,
-                                            HealChannelPriority::High,
-                                        ),
-                                    )
-                                    .await?;
-                                    found_objects = true;
+                            for candidate in discovery.candidates {
+                                let sub_quorum_candidate = candidate.replica_count < disks_quorum;
+                                let version_id = candidate.validated_version().map(|id| id.to_string());
+                                let identity = (candidate.object.clone(), version_id.clone(), candidate.kind.clone());
+                                if seen_heal_candidates.len() >= MAX_META_CACHE_HEAL_CANDIDATES
+                                    && !seen_heal_candidates.contains(&identity)
+                                {
                                     continue;
                                 }
-                            };
-
-                            for fiv in fivs.versions {
-                                let version_id = fiv.version_id.and_then(|v| if v.is_nil() { None } else { Some(v.to_string()) });
-                                self.send_required_scanner_heal_request(
-                                    PendingScannerHealKind::Object,
-                                    bucket.clone(),
-                                    Some(entry.name.clone()),
-                                    version_id.clone(),
-                                    build_object_heal_request(
+                                if !seen_heal_candidates.insert(identity) {
+                                    continue;
+                                }
+                                let request = if candidate.is_unversioned() {
+                                    build_non_destructive_object_heal_request(
                                         bucket.clone(),
-                                        entry.name.clone(),
-                                        version_id,
+                                        candidate.object.clone(),
                                         self.scan_mode,
                                         HealChannelPriority::High,
-                                    ),
+                                    )
+                                } else {
+                                    build_object_heal_request(
+                                        bucket.clone(),
+                                        candidate.object.clone(),
+                                        version_id.clone(),
+                                        self.scan_mode,
+                                        HealChannelPriority::High,
+                                    )
+                                };
+                                (self.update_current_path)(&candidate.object).await;
+                                let admission = self.send_required_scanner_heal_request(
+                                    PendingScannerHealKind::Object,
+                                    bucket.clone(),
+                                    Some(candidate.object.clone()),
+                                    version_id.clone(),
+                                    request,
                                 )
                                 .await?;
+                                if admission.is_admitted() {
+                                    counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
+                                } else if sub_quorum_candidate {
+                                    self.mark_pending_scanner_heal_reason(
+                                        PendingScannerHealKind::Object,
+                                        &bucket,
+                                        Some(&candidate.object),
+                                        version_id.as_deref(),
+                                        "sub_quorum_metadata",
+                                    );
+                                }
+                                found_objects = true;
+                            }
+
+                            // Candidates beyond the main cap remain exact
+                            // version requests; never downgrade them to a
+                            // latest-version (version_id=None) heal.
+                            for candidate in discovery.truncated_candidates {
+                                let version_id = candidate.validated_version().map(|id| id.to_string());
+                                let identity = (candidate.object.clone(), version_id.clone(), candidate.kind.clone());
+                                if seen_truncated_objects.len() >= MAX_META_CACHE_HEAL_TRUNCATED_OBJECTS
+                                    && !seen_truncated_objects.contains(&candidate.object)
+                                {
+                                    continue;
+                                }
+                                seen_truncated_objects.insert(candidate.object.clone());
+                                if !seen_heal_candidates.insert(identity) {
+                                    continue;
+                                }
+                                let request = build_object_heal_request(
+                                    bucket.clone(),
+                                    candidate.object.clone(),
+                                    version_id.clone(),
+                                    self.scan_mode,
+                                    HealChannelPriority::High,
+                                );
+                                (self.update_current_path)(&candidate.object).await;
+                                let admission = self
+                                    .send_required_scanner_heal_request(
+                                        PendingScannerHealKind::Object,
+                                        bucket.clone(),
+                                        Some(candidate.object.clone()),
+                                        version_id,
+                                        request,
+                                    )
+                                    .await?;
+                                if admission.is_admitted() {
+                                    counter!(METRIC_SCANNER_HEAL_DISCOVERY_QUEUED_TOTAL).increment(1);
+                                }
                                 found_objects = true;
                             }
 

crates/scanner/src/scanner_folder/item_actions.rs

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

crates/scanner/src/scanner_folder/ledger.rs

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

crates/scanner/src/scanner_folder/tests.rs

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

crates/scanner/src/scanner_io.rs

@@ -18,8 +18,8 @@ use crate::scanner_folder::{ScannerItem, scan_data_folder};
 use crate::sleeper::SCANNER_SLEEPER;
 use crate::{
     DATA_USAGE_CACHE_NAME, DATA_USAGE_ROOT, DataUsageCache, DataUsageCacheInfo, DataUsageCachePrepareOutcome,
-    DataUsageCacheSource, DataUsageEntry, DataUsageEntryInfo, DataUsageInfo, DataUsageScanPlanDigest, DataUsageSnapshotSetState,
-    ScannerError, SizeSummary, TierStats,
+    DataUsageCacheSource, DataUsageEntry, DataUsageEntryInfo, DataUsageInfo, DataUsageScanPlanDigest, ScannerError, SizeSummary,
+    TierStats,
 };
 use futures::future::join_all;
 use metrics::counter;
@@ -278,17 +278,6 @@ async fn publish_usage_snapshot(
     Ok(true)
 }
 
-async fn publish_observational_snapshot(
-    updates: &mpsc::Sender<DataUsageInfo>,
-    mut data_usage_info: DataUsageInfo,
-) -> Result<bool> {
-    data_usage_info.usage_snapshot_complete = false;
-    data_usage_info.usage_snapshot_partial = true;
-    data_usage_info.usage_snapshot_converged = Some(false);
-    send_data_usage_update(updates, data_usage_info).await?;
-    Ok(true)
-}
-
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 enum ScannerCycleActivityStatus {
     Unchanged,

crates/scanner/src/scanner_io/cache.rs

@@ -188,7 +188,7 @@ pub(super) fn completed_data_usage_info(
     }
 
     let mut total = DataUsageEntry::default();
-    let mut bucket_entries = HashMap::with_capacity(all_buckets.len());
+    let mut buckets_usage = HashMap::with_capacity(all_buckets.len());
     for bucket in all_buckets {
         let mut merged = DataUsageEntry::default();
         for result in results {
@@ -200,14 +200,10 @@ pub(super) fn completed_data_usage_info(
         if !total.checked_merge(&merged) {
             return None;
         }
-        bucket_entries.insert(bucket.clone(), merged);
+        buckets_usage.insert(bucket.clone(), checked_bucket_usage_info(&merged)?);
     }
 
     let merged_last_update = results.iter().filter_map(|result| result.info.last_update).max()?;
-    let buckets_usage = bucket_entries
-        .iter()
-        .map(|(bucket, entry)| Some((bucket.clone(), checked_bucket_usage_info(entry)?)))
-        .collect::<Option<HashMap<_, _>>>()?;
     let bucket_sizes = buckets_usage
         .iter()
         .map(|(bucket, usage)| (bucket.clone(), usage.size))
@@ -229,145 +225,6 @@ pub(super) fn completed_data_usage_info(
     Some((data_usage_info, merged_last_update))
 }
 
-/// Build a non-authoritative view from the set snapshots that completed this
-/// cycle plus compatible per-set last-known-good caches.  The caller must
-/// persist this result only on the observational object; a missing set is
-/// intentionally represented by an incomplete state and is never treated as
-/// an empty set.
-pub(super) fn observational_data_usage_info(
-    results: &[DataUsageCache],
-    expected_sources: &HashSet<DataUsageCacheSource>,
-    all_buckets: &[String],
-    expected_plan_digest: DataUsageScanPlanDigest,
-    scanner_cycle: u64,
-    leader_epoch: u64,
-) -> Option<(DataUsageInfo, SystemTime)> {
-    let mut by_source = HashMap::with_capacity(results.len());
-    for result in results {
-        let source = result.info.source?;
-        if !expected_sources.contains(&source) || by_source.insert(source, result).is_some() {
-            return None;
-        }
-    }
-
-    let mut usable = Vec::new();
-    let mut set_states = Vec::with_capacity(expected_sources.len());
-    let mut sources = expected_sources.iter().copied().collect::<Vec<_>>();
-    sources.sort_by_key(|source| (source.pool_index, source.set_index));
-    for source in sources {
-        let result = by_source.get(&source).copied();
-        let current = result.filter(|result| {
-            result.info.snapshot_complete
-                && result.info.next_cycle == scanner_cycle
-                && result.info.leader_epoch == leader_epoch
-                && result.info.scan_plan_digest == Some(expected_plan_digest)
-        });
-        let lkg = result.filter(|result| {
-            !result.info.snapshot_complete
-                && result.info.lkg_snapshot_complete
-                && result.info.lkg_scan_plan_digest == Some(expected_plan_digest)
-                && result.info.lkg_leader_epoch.is_some_and(|epoch| {
-                    epoch < leader_epoch
-                        || (epoch == leader_epoch && result.info.lkg_next_cycle.is_some_and(|cycle| cycle <= scanner_cycle))
-                })
-        });
-        let current_snapshot = current.is_some();
-        let selected = current.or(lkg);
-        if let Some(selected) = selected {
-            let (cycle, epoch, digest, last_update, complete) = if current_snapshot {
-                (
-                    Some(selected.info.next_cycle),
-                    Some(selected.info.leader_epoch),
-                    selected.info.scan_plan_digest.map(|digest| digest.0),
-                    selected.info.last_update,
-                    true,
-                )
-            } else {
-                (
-                    selected.info.lkg_next_cycle,
-                    selected.info.lkg_leader_epoch,
-                    selected.info.lkg_scan_plan_digest.map(|digest| digest.0),
-                    selected.info.lkg_last_update,
-                    false,
-                )
-            };
-            set_states.push(DataUsageSnapshotSetState {
-                pool_index: u64::try_from(source.pool_index).ok()?,
-                set_index: u64::try_from(source.set_index).ok()?,
-                scanner_cycle: cycle,
-                scanner_epoch: epoch,
-                scan_plan_digest: digest,
-                complete,
-                tombstone: false,
-            });
-            usable.push((selected, last_update));
-        } else {
-            set_states.push(DataUsageSnapshotSetState {
-                pool_index: u64::try_from(source.pool_index).ok()?,
-                set_index: u64::try_from(source.set_index).ok()?,
-                scanner_cycle: None,
-                scanner_epoch: None,
-                scan_plan_digest: Some(expected_plan_digest.0),
-                complete: false,
-                tombstone: false,
-            });
-        }
-    }
-    if usable.is_empty() {
-        return None;
-    }
-
-    let mut total = DataUsageEntry::default();
-    let mut bucket_entries = HashMap::with_capacity(all_buckets.len());
-    let mut merged_last_update = None;
-    for (result, last_update) in usable {
-        if let Some(update) = last_update {
-            merged_last_update = Some(merged_last_update.map_or(update, |current: SystemTime| current.max(update)));
-        }
-        for bucket in all_buckets {
-            let Some(entry) = result.checked_flatten(bucket) else {
-                continue;
-            };
-            let bucket_entry = bucket_entries.entry(bucket.clone()).or_insert_with(DataUsageEntry::default);
-            if !bucket_entry.checked_merge(&entry) {
-                return None;
-            }
-            if !total.checked_merge(&entry) {
-                return None;
-            }
-        }
-    }
-    let merged_last_update = merged_last_update?;
-    let buckets_usage = bucket_entries
-        .iter()
-        .map(|(bucket, entry)| Some((bucket.clone(), checked_bucket_usage_info(entry)?)))
-        .collect::<Option<HashMap<_, _>>>()?;
-    Some((
-        DataUsageInfo {
-            last_update: Some(merged_last_update),
-            scanner_cycle: Some(scanner_cycle),
-            scanner_epoch: Some(leader_epoch),
-            objects_total_count: u64::try_from(total.objects).ok()?,
-            versions_total_count: u64::try_from(total.versions).ok()?,
-            delete_markers_total_count: u64::try_from(total.delete_markers).ok()?,
-            objects_total_size: u64::try_from(total.size).ok()?,
-            tier_stats: total.all_tier_stats.filter(|tiers| !tiers.is_empty()),
-            buckets_count: u64::try_from(buckets_usage.len()).ok()?,
-            bucket_sizes: buckets_usage
-                .iter()
-                .map(|(bucket, usage)| (bucket.clone(), usage.size))
-                .collect(),
-            buckets_usage,
-            usage_snapshot_complete: false,
-            usage_snapshot_partial: true,
-            usage_snapshot_converged: Some(false),
-            usage_snapshot_set_states: set_states,
-            ..Default::default()
-        },
-        merged_last_update,
-    ))
-}
-
 pub(super) async fn send_cache_root_entry_info(
     bucket_result_tx: &mpsc::Sender<DataUsageEntryInfo>,
     cache: &DataUsageCache,

crates/scanner/src/scanner_io/io_cache.rs

@@ -40,21 +40,6 @@ impl ScannerIOCache for SetDisks {
         let set_label = self.set_index.to_string();
 
         let source = DataUsageCacheSource::new(self.pool_index, self.set_index);
-        let mut old_cache = DataUsageCache::default();
-        if let Err(e) = old_cache.load(self.clone(), DATA_USAGE_CACHE_NAME).await {
-            warn!(
-                target: "rustfs::scanner::io",
-                event = EVENT_SCANNER_CACHE_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_IO,
-                pool = self.pool_index,
-                set = self.set_index,
-                cache_name = DATA_USAGE_CACHE_NAME,
-                state = "old_cache_load_failed",
-                error = %e,
-                "Scanner old data usage cache load failed; rebuilding from bucket caches"
-            );
-        }
         if buckets.is_empty() {
             let now = SystemTime::now();
             let mut cache = DataUsageCache {
@@ -95,24 +80,6 @@ impl ScannerIOCache for SetDisks {
                 "Scanner set state found no online disks"
             );
             reset_disk_bucket_scan_gauges(&pool_label, &set_label);
-            let lkg = old_cache.info.snapshot_complete.then(|| old_cache.clone());
-            let mut incomplete_scope = lkg.clone().unwrap_or_default();
-            incomplete_scope.info.name = DATA_USAGE_ROOT.to_string();
-            incomplete_scope.info.next_cycle = want_cycle;
-            incomplete_scope.info.last_update = None;
-            incomplete_scope.info.leader_epoch = leader_epoch;
-            incomplete_scope.info.source = Some(source);
-            incomplete_scope.info.snapshot_complete = false;
-            incomplete_scope.info.scan_plan_digest = Some(scan_plan_digest);
-            incomplete_scope.info.cache_key_format = DATA_USAGE_CACHE_KEY_FORMAT;
-            if let Some(lkg) = lkg {
-                incomplete_scope.info.lkg_snapshot_complete = true;
-                incomplete_scope.info.lkg_next_cycle = Some(lkg.info.next_cycle);
-                incomplete_scope.info.lkg_last_update = lkg.info.last_update;
-                incomplete_scope.info.lkg_leader_epoch = Some(lkg.info.leader_epoch);
-                incomplete_scope.info.lkg_scan_plan_digest = lkg.info.scan_plan_digest;
-            }
-            let _ = updates.send(incomplete_scope).await;
             return Ok(());
         }
         // Preserve the original set topology across capability filtering. During
@@ -195,24 +162,6 @@ impl ScannerIOCache for SetDisks {
                 "Scanner set state found no usable namespace scanner disks"
             );
             reset_disk_bucket_scan_gauges(&pool_label, &set_label);
-            let lkg = old_cache.info.snapshot_complete.then(|| old_cache.clone());
-            let mut incomplete_scope = lkg.clone().unwrap_or_default();
-            incomplete_scope.info.name = DATA_USAGE_ROOT.to_string();
-            incomplete_scope.info.next_cycle = want_cycle;
-            incomplete_scope.info.last_update = None;
-            incomplete_scope.info.leader_epoch = leader_epoch;
-            incomplete_scope.info.source = Some(source);
-            incomplete_scope.info.snapshot_complete = false;
-            incomplete_scope.info.scan_plan_digest = Some(scan_plan_digest);
-            incomplete_scope.info.cache_key_format = DATA_USAGE_CACHE_KEY_FORMAT;
-            if let Some(lkg) = lkg {
-                incomplete_scope.info.lkg_snapshot_complete = true;
-                incomplete_scope.info.lkg_next_cycle = Some(lkg.info.next_cycle);
-                incomplete_scope.info.lkg_last_update = lkg.info.last_update;
-                incomplete_scope.info.lkg_leader_epoch = Some(lkg.info.leader_epoch);
-                incomplete_scope.info.lkg_scan_plan_digest = lkg.info.scan_plan_digest;
-            }
-            let _ = updates.send(incomplete_scope).await;
             return Ok(());
         }
         let set_disk_inventory = Arc::new(scanner_set_disk_inventory(self.as_ref()).await);
@@ -254,15 +203,22 @@ impl ScannerIOCache for SetDisks {
         record_disk_bucket_scans_active(0, &pool_label, &set_label);
         let _reset_disk_bucket_scan_gauges = DiskBucketScanGaugeReset::new(pool_label.clone(), set_label.clone());
 
-        let old_lkg = old_cache.info.snapshot_complete.then(|| {
-            (
-                old_cache.info.next_cycle,
-                old_cache.info.last_update,
-                old_cache.info.leader_epoch,
-                old_cache.info.scan_plan_digest,
-            )
-        });
-        let prepare_outcome = match old_cache.prepare_for_scan(
+        let mut old_cache = DataUsageCache::default();
+        if let Err(e) = old_cache.load(self.clone(), DATA_USAGE_CACHE_NAME).await {
+            warn!(
+                target: "rustfs::scanner::io",
+                event = EVENT_SCANNER_CACHE_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_IO,
+                pool = self.pool_index,
+                set = self.set_index,
+                cache_name = DATA_USAGE_CACHE_NAME,
+                state = "old_cache_load_failed",
+                error = %e,
+                "Scanner old data usage cache load failed; rebuilding from bucket caches"
+            );
+        }
+        match old_cache.prepare_for_scan(
             DATA_USAGE_ROOT,
             want_cycle,
             leader_epoch,
@@ -303,16 +259,7 @@ impl ScannerIOCache for SetDisks {
                 );
                 return Ok(());
             }
-            outcome => outcome,
-        };
-        if matches!(prepare_outcome, DataUsageCachePrepareOutcome::Reused)
-            && let Some((cycle, last_update, epoch, digest)) = old_lkg
-        {
-            old_cache.info.lkg_snapshot_complete = true;
-            old_cache.info.lkg_next_cycle = Some(cycle);
-            old_cache.info.lkg_last_update = last_update;
-            old_cache.info.lkg_leader_epoch = Some(epoch);
-            old_cache.info.lkg_scan_plan_digest = digest;
+            DataUsageCachePrepareOutcome::Reused | DataUsageCachePrepareOutcome::Reset => {}
         }
 
         let mut cache = DataUsageCache {
@@ -1152,29 +1099,23 @@ impl ScannerIOCache for SetDisks {
                 cache.info.next_cycle = want_cycle;
                 cache.info.last_update.get_or_insert_with(SystemTime::now);
                 cache.info.snapshot_complete = true;
-                cache.info.lkg_snapshot_complete = false;
-                cache.info.lkg_next_cycle = None;
-                cache.info.lkg_last_update = None;
-                cache.info.lkg_leader_epoch = None;
-                cache.info.lkg_scan_plan_digest = None;
                 cache.clone()
             };
             let _ = persist_and_publish_cache_snapshot(self.clone(), &updates, cache_snapshot, cache_cycle_floor.as_ref()).await;
         } else {
-            let mut incomplete_scope = cache_mutex.lock().await.clone();
-            incomplete_scope.info.name = DATA_USAGE_ROOT.to_string();
-            incomplete_scope.info.next_cycle = want_cycle;
-            incomplete_scope.info.last_update = None;
-            incomplete_scope.info.leader_epoch = leader_epoch;
-            incomplete_scope.info.source = Some(source);
-            incomplete_scope.info.snapshot_complete = false;
-            incomplete_scope.info.scan_plan_digest = Some(scan_plan_digest);
-            incomplete_scope.info.cache_key_format = DATA_USAGE_CACHE_KEY_FORMAT;
-            incomplete_scope.info.lkg_snapshot_complete = old_cache.info.lkg_snapshot_complete;
-            incomplete_scope.info.lkg_next_cycle = old_cache.info.lkg_next_cycle;
-            incomplete_scope.info.lkg_last_update = old_cache.info.lkg_last_update;
-            incomplete_scope.info.lkg_leader_epoch = old_cache.info.lkg_leader_epoch;
-            incomplete_scope.info.lkg_scan_plan_digest = old_cache.info.lkg_scan_plan_digest;
+            let incomplete_scope = DataUsageCache {
+                info: DataUsageCacheInfo {
+                    name: DATA_USAGE_ROOT.to_string(),
+                    next_cycle: want_cycle,
+                    leader_epoch,
+                    source: Some(source),
+                    snapshot_complete: false,
+                    scan_plan_digest: Some(scan_plan_digest),
+                    cache_key_format: DATA_USAGE_CACHE_KEY_FORMAT,
+                    ..Default::default()
+                },
+                cache: HashMap::new(),
+            };
             if let Err(e) = updates.send(incomplete_scope).await {
                 error!(
                     target: "rustfs::scanner::io",

crates/scanner/src/scanner_io/io_cycle.rs

@@ -234,7 +234,6 @@ impl ScannerIOCycle for ECStore {
                 let active_set_scans_clone = active_set_scans.clone();
 
                 let (tx, mut rx) = mpsc::channel::<DataUsageCache>(1);
-                let failed_scope_tx = tx.clone();
 
                 // Spawn task to receive and store results
                 let receiver_fut = tokio::spawn(async move {
@@ -315,21 +314,6 @@ impl ScannerIOCycle for ECStore {
                             state = "set_scan_failed",
                             "Scanner set scan failed; continuing cycle"
                         );
-                        let _ = failed_scope_tx
-                            .send(DataUsageCache {
-                                info: DataUsageCacheInfo {
-                                    name: DATA_USAGE_ROOT.to_string(),
-                                    next_cycle: want_cycle_clone,
-                                    leader_epoch,
-                                    source: Some(source),
-                                    snapshot_complete: false,
-                                    scan_plan_digest: Some(scan_plan_digest),
-                                    cache_key_format: DATA_USAGE_CACHE_KEY_FORMAT,
-                                    ..Default::default()
-                                },
-                                cache: HashMap::new(),
-                            })
-                            .await;
                         let mut first_err = first_err_mutex_clone.lock().await;
                         record_set_scan_failure(&mut first_err, e);
                     }
@@ -386,19 +370,6 @@ impl ScannerIOCycle for ECStore {
             budget_elapsed,
             ctx.is_cancelled(),
         );
-        let observational_usage = completed_usage
-            .is_none()
-            .then(|| {
-                observational_data_usage_info(
-                    &results,
-                    &expected_sources,
-                    &all_bucket_names,
-                    scan_plan_digest,
-                    want_cycle,
-                    leader_epoch,
-                )
-            })
-            .flatten();
         let structurally_complete_snapshot = result.is_ok() && completed_all_sets && completed_usage.is_some();
         let cycle_status = classify_nsscanner_cycle(
             structurally_complete_snapshot,
@@ -410,10 +381,6 @@ impl ScannerIOCycle for ECStore {
         );
         if let Some((data_usage_info, _)) = completed_usage {
             publish_usage_snapshot(&updates, cycle_status, data_usage_info).await?;
-        } else if !ctx.is_cancelled()
-            && let Some((data_usage_info, _)) = observational_usage
-        {
-            publish_observational_snapshot(&updates, data_usage_info).await?;
         }
         let dirty_usage_clear = should_clear_dirty_usage_snapshot(
             result.is_ok(),

crates/scanner/src/scanner_io/publish_gate_tests.rs

@@ -105,160 +105,6 @@ fn completed_data_usage_info_for_test(
     completed_data_usage_info(results, &expected_sources, all_buckets, true, budget_elapsed, cancelled)
 }
 
-fn lkg_root_cache(bucket: &str, objects: usize, source: DataUsageCacheSource) -> DataUsageCache {
-    let mut cache = completed_root_cache(bucket, objects, 10, source);
-    cache.info.snapshot_complete = false;
-    cache.info.next_cycle = 8;
-    cache.info.leader_epoch = 3;
-    cache.info.lkg_snapshot_complete = true;
-    cache.info.lkg_next_cycle = Some(7);
-    cache.info.lkg_last_update = cache.info.last_update;
-    cache.info.lkg_leader_epoch = Some(3);
-    cache.info.lkg_scan_plan_digest = Some(TEST_PLAN_DIGEST);
-    cache
-}
-
-#[test]
-fn partial_usage_is_observational_not_authoritative_for_quota() {
-    let all_buckets = vec!["bucket".to_string()];
-    let current_source = DataUsageCacheSource::new(0, 0);
-    let stalled_source = DataUsageCacheSource::new(1, 0);
-    let mut current = completed_root_cache("bucket", 2, 20, current_source);
-    current.info.next_cycle = 8;
-    current.info.leader_epoch = 3;
-    let stalled = lkg_root_cache("bucket", 1, stalled_source);
-    let expected = HashSet::from([current_source, stalled_source]);
-
-    assert!(
-        completed_data_usage_info(&[current.clone(), stalled.clone()], &expected, &all_buckets, true, false, false).is_none()
-    );
-    let (observed, _) = observational_data_usage_info(&[current, stalled], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
-        .expect("a completed set should produce an observational view");
-    assert!(observed.usage_snapshot_partial);
-    assert!(!observed.usage_snapshot_complete);
-    assert_eq!(observed.usage_snapshot_converged, Some(false));
-    assert_eq!(observed.usage_snapshot_set_states.len(), 2);
-}
-
-#[test]
-fn lkg_scope_does_not_count_as_current_cycle_completion() {
-    let source = DataUsageCacheSource::new(0, 0);
-    let mut lkg = lkg_root_cache("bucket", 1, source);
-    lkg.info.last_update = None;
-    let expected = HashSet::from([source]);
-    assert!(!scanner_results_form_complete_snapshot(&[lkg], &expected));
-}
-
-#[test]
-fn stale_quota_uses_complete_baseline_plus_positive_deltas() {
-    let all_buckets = vec!["bucket".to_string()];
-    let source = DataUsageCacheSource::new(0, 0);
-    let mut current = completed_root_cache("bucket", 3, 20, source);
-    current.info.next_cycle = 8;
-    current.info.leader_epoch = 3;
-    let expected = HashSet::from([source]);
-    let (observed, _) = observational_data_usage_info(&[current], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
-        .expect("complete set data is a valid observational baseline");
-    assert_eq!(observed.objects_total_size, 30);
-    assert_eq!(observed.usage_snapshot_set_states[0].complete, true);
-}
-
-#[test]
-fn negative_delta_waits_for_set_reconciliation() {
-    let all_buckets = vec!["bucket".to_string()];
-    let source = DataUsageCacheSource::new(0, 0);
-    let mut stalled = lkg_root_cache("bucket", 4, source);
-    stalled.info.lkg_scan_plan_digest = Some(DataUsageScanPlanDigest([9; 32]));
-    let expected = HashSet::from([source]);
-    assert!(observational_data_usage_info(&[stalled], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3).is_none());
-}
-
-#[test]
-fn set_membership_add_remove_uses_generation_and_tombstone() {
-    let state = DataUsageSnapshotSetState {
-        pool_index: 1,
-        set_index: 2,
-        scanner_cycle: Some(9),
-        scanner_epoch: Some(4),
-        scan_plan_digest: Some(TEST_PLAN_DIGEST.0),
-        complete: false,
-        tombstone: true,
-    };
-    let encoded = serde_json::to_vec(&state).expect("set state should serialize");
-    let decoded: DataUsageSnapshotSetState = serde_json::from_slice(&encoded).expect("set state should deserialize");
-    assert_eq!(decoded, state);
-
-    let snapshot = DataUsageInfo {
-        last_update: Some(SystemTime::UNIX_EPOCH + Duration::from_secs(10)),
-        scanner_cycle: Some(9),
-        scanner_epoch: Some(4),
-        buckets_count: 0,
-        usage_snapshot_converged: Some(false),
-        usage_snapshot_partial: true,
-        usage_snapshot_set_states: vec![
-            DataUsageSnapshotSetState {
-                pool_index: 0,
-                set_index: 0,
-                scanner_cycle: Some(9),
-                scanner_epoch: Some(4),
-                scan_plan_digest: Some(TEST_PLAN_DIGEST.0),
-                complete: true,
-                tombstone: false,
-            },
-            state,
-        ],
-        ..Default::default()
-    };
-    assert!(snapshot.is_valid_partial_snapshot());
-}
-
-#[test]
-fn old_set_completion_cannot_overwrite_new_aggregate() {
-    let all_buckets = vec!["bucket".to_string()];
-    let source = DataUsageCacheSource::new(0, 0);
-    let mut old = completed_root_cache("bucket", 1, 20, source);
-    old.info.next_cycle = 7;
-    old.info.leader_epoch = 2;
-    let expected = HashSet::from([source]);
-    assert!(observational_data_usage_info(&[old], &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3).is_none());
-}
-
-#[test]
-fn usage_aggregate_survives_restart_and_leader_failover() {
-    let all_buckets = vec!["bucket".to_string()];
-    let source = DataUsageCacheSource::new(0, 0);
-    let mut lkg = lkg_root_cache("bucket", 5, source);
-    lkg.info.lkg_leader_epoch = Some(4);
-    lkg.info.lkg_next_cycle = Some(9);
-    let expected = HashSet::from([source]);
-    let (observed, _) = observational_data_usage_info(&[lkg], &expected, &all_buckets, TEST_PLAN_DIGEST, 10, 5)
-        .expect("compatible LKG should survive a leader change");
-    assert_eq!(observed.usage_snapshot_set_states[0].scanner_epoch, Some(4));
-    assert_eq!(observed.objects_total_size, 50);
-}
-
-#[test]
-fn usage_aggregate_cost_is_linear_in_set_count() {
-    let all_buckets = vec!["bucket".to_string()];
-    let mut results = Vec::new();
-    let mut expected = HashSet::new();
-    for index in 0..32 {
-        let source = DataUsageCacheSource::new(index, 0);
-        expected.insert(source);
-        let mut cache = completed_root_cache("bucket", 1, 20, source);
-        cache.info.next_cycle = 8;
-        cache.info.leader_epoch = 3;
-        results.push(cache);
-    }
-    let (observed, _) = observational_data_usage_info(&results, &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
-        .expect("all set snapshots should aggregate");
-    assert_eq!(observed.objects_total_count, 32);
-    let reversed = results.iter().rev().cloned().collect::<Vec<_>>();
-    let (reversed_observed, _) = observational_data_usage_info(&reversed, &expected, &all_buckets, TEST_PLAN_DIGEST, 8, 3)
-        .expect("reordered set snapshots should aggregate");
-    assert_eq!(observed.usage_snapshot_set_states, reversed_observed.usage_snapshot_set_states);
-}
-
 #[test]
 fn completed_data_usage_info_publishes_tier_stats_across_sets() {
     let all_buckets = vec!["bucket-a".to_string(), "bucket-b".to_string()];

deny.toml

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

docs/testing/ci-gates.md

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

rustfs/Cargo.toml

@@ -336,13 +336,13 @@ opentelemetry = { workspace = true }
 tracing-opentelemetry = { workspace = true }
 # Data structures
 hashbrown = { workspace = true, features = ["serde", "rayon"] }
-mimalloc = { workspace = true }
+rustfs-mimalloc = { workspace = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 libsystemd.workspace = true
 
 [target.'cfg(not(target_os = "windows"))'.dependencies]
-libmimalloc-sys.workspace = true
+rustfs-mimalloc-sys.workspace = true
 
 [dev-dependencies]
 uuid = { workspace = true, features = ["v4", "v5", "fast-rng", "macro-diagnostics"] }

rustfs/src/allocator_reclaim.rs

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

rustfs/src/main.rs

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

rustfs/src/memory_observability.rs

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