Merge branch 'main' into cxymds/fix-1927-durable-checkpoint

Signed-off-by: houseme <housemecn@gmail.com>
docs(testing): document CI gate matrix (#6412 )
2026-08-23 04:39:04 +00:00 · 2026-08-23 12:27:11 +08:00 · 2026-08-23 12:09:06 +08:00 · 2026-08-23 12:07:25 +08:00 · 2026-08-23 12:07:20 +08:00 · 2026-08-23 07:58:22 +08:00
32 changed files with 1188 additions and 1034 deletions
@@ -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):
@@ -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

@@ -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",
@@ -9587,6 +9564,7 @@ dependencies = [
 "serde",
 "serde_json",
 "serial_test",
+ "sha2 0.11.0",
 "temp-env",
 "tempfile",
 "thiserror 2.0.20",
@@ -9875,6 +9853,24 @@ 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"
@@ -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" }
@@ -23,32 +23,21 @@
 //! unconsumed intents is the consumer's job (see `rustfs-heal`
 //! `heal::mrf_queue`), mirroring MinIO's `.heal/mrf/list.bin`.

-use std::collections::HashMap;
-use std::collections::hash_map::RandomState;
-use std::hash::{BuildHasher, Hash};
-use std::sync::atomic::AtomicU64;
-use std::sync::atomic::AtomicUsize;
 use std::sync::{
-    Arc, Mutex, OnceLock,
+    Arc, OnceLock,
    atomic::{AtomicBool, Ordering},
 };
-use std::time::{Duration, Instant};
 use tokio::sync::mpsc;
 use uuid::Uuid;

 /// Bounded capacity of the global MRF channel. Backpressure is resolved by
 /// dropping (and counting) intents, never by blocking the producer.
 const MRF_CHANNEL_CAPACITY: usize = 8192;
-const MRF_COALESCER_SHARDS: usize = 16;
-const MRF_COALESCER_MAX_KEYS: usize = 8192;
-const MRF_COALESCER_MAX_BYTES: usize = 16 * 1024 * 1024;
-const MRF_COALESCER_TTL: Duration = Duration::from_secs(60);
-const MRF_MAX_IDENTITY_COMPONENT: usize = 1024;

 /// Why an intent was produced. Drives the heal priority mapping on the
 /// consumer side (DecodeFailure -> Urgent, MetadataCorruption -> High,
 /// PartialWrite -> Normal).
-#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum MrfKind {
    /// Erasure decode failed while serving a read (read path).
    DecodeFailure,
@@ -78,52 +67,12 @@ pub struct MrfIntent {
    /// Version the intent targets, as raw UUID bytes.
    pub version_id: Option<[u8; 16]>,
    pub kind: MrfKind,
-    /// Stable erasure-set scope when the producer has it. Kept optional so
-    /// metadata corruption and legacy producers do not invent a scope.
-    pub scope: Option<MrfScope>,
-    /// Generation of the node-local ingress lease. It is not persisted in
-    /// the journal; replayed records acquire a fresh lease when re-enqueued.
-    pub lease: Option<MrfIngressLease>,
    pub enqueued_at_ms: u64,
    /// Times this intent has already been offered to the heal manager.
    /// Dropped by the consumer once it reaches `MRF_MAX_ATTEMPTS`.
    pub attempts: u8,
 }

-#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
-pub struct MrfScope {
-    pub pool_index: u32,
-    pub set_index: u32,
-}
-
-/// Opaque generation used to release exactly the admission that created an
-/// ingress entry. A generation prevents a late terminal callback from
-/// deleting a newer retry for the same identity (ABA).
-#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
-pub struct MrfIngressLease(u64);
-
-impl MrfIngressLease {
-    const fn new(value: u64) -> Self {
-        Self(value)
-    }
-}
-
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
-pub enum MrfDropReason {
-    Disabled,
-    Uninitialized,
-    Full,
-    OversizedIdentity,
-    CoalescerFull,
-}
-
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
-pub enum MrfIngressResult {
-    Enqueued,
-    Coalesced,
-    Dropped(MrfDropReason),
-}
-
 /// Consumer-side retry ceiling before an intent is given up on.
 pub const MRF_MAX_ATTEMPTS: u8 = 3;

@@ -138,159 +87,6 @@ impl MrfIntent {

 static GLOBAL_MRF_SENDER: OnceLock<mpsc::Sender<MrfIntent>> = OnceLock::new();

-#[derive(Clone, Debug, PartialEq, Eq, Hash)]
-struct MrfIdentityKey {
-    kind: MrfKind,
-    bucket: Arc<str>,
-    object: Arc<str>,
-    version_id: Option<[u8; 16]>,
-    scope: Option<MrfScope>,
-}
-
-#[derive(Debug)]
-struct IngressEntry {
-    lease: MrfIngressLease,
-    expires_at: Instant,
-    bytes: usize,
-}
-
-type MrfCoalescerShard = Mutex<HashMap<MrfIdentityKey, IngressEntry>>;
-type MrfCoalescer = Box<[MrfCoalescerShard]>;
-
-static MRF_COALESCER: OnceLock<MrfCoalescer> = OnceLock::new();
-static NEXT_MRF_LEASE: AtomicU64 = AtomicU64::new(1);
-static MRF_COALESCER_COUNT: AtomicUsize = AtomicUsize::new(0);
-static MRF_COALESCER_BYTES: AtomicUsize = AtomicUsize::new(0);
-static MRF_HASH_STATE: OnceLock<RandomState> = OnceLock::new();
-
-fn coalescer() -> &'static [MrfCoalescerShard] {
-    MRF_COALESCER.get_or_init(|| {
-        (0..MRF_COALESCER_SHARDS)
-            .map(|_| Mutex::new(HashMap::new()))
-            .collect::<Vec<_>>()
-            .into_boxed_slice()
-    })
-}
-
-fn key_shard(key: &MrfIdentityKey) -> usize {
-    let hash = MRF_HASH_STATE.get_or_init(RandomState::new).hash_one(key);
-    usize::try_from(hash).unwrap_or(0) % MRF_COALESCER_SHARDS
-}
-
-fn canonical_version(version_id: Option<Uuid>) -> Option<[u8; 16]> {
-    version_id
-        .filter(|version| !version.is_nil())
-        .map(|version| *version.as_bytes())
-}
-
-fn canonical_identity(
-    kind: MrfKind,
-    version_id: Option<[u8; 16]>,
-    scope: Option<MrfScope>,
-) -> (Option<[u8; 16]>, Option<MrfScope>) {
-    let version_id = version_id.filter(|bytes| *bytes != [0; 16]);
-    match kind {
-        MrfKind::MetadataCorruption => (None, None),
-        MrfKind::DecodeFailure | MrfKind::PartialWrite => (version_id, scope),
-    }
-}
-
-fn identity_estimated_bytes(key: &MrfIdentityKey) -> usize {
-    64usize
-        .saturating_add(key.bucket.len())
-        .saturating_add(key.object.len())
-        .saturating_add(key.version_id.map_or(0, |_| 16))
-        .saturating_add(key.scope.map_or(0, |_| 8))
-}
-
-fn reserve(counter: &AtomicUsize, limit: usize, amount: usize) -> bool {
-    let mut current = counter.load(Ordering::Relaxed);
-    loop {
-        let Some(next) = current.checked_add(amount) else {
-            return false;
-        };
-        if next > limit {
-            return false;
-        }
-        match counter.compare_exchange_weak(current, next, Ordering::Relaxed, Ordering::Relaxed) {
-            Ok(_) => return true,
-            Err(observed) => current = observed,
-        }
-    }
-}
-
-fn coalescer_admit(key: MrfIdentityKey) -> Result<MrfIngressLease, MrfIngressResult> {
-    let shard = key_shard(&key);
-    let mut entries = coalescer()[shard]
-        .lock()
-        .map_err(|_| MrfIngressResult::Dropped(MrfDropReason::CoalescerFull))?;
-    let now = Instant::now();
-    let before = entries.len();
-    let mut expired_bytes = 0usize;
-    entries.retain(|_, entry| {
-        if entry.expires_at > now {
-            true
-        } else {
-            expired_bytes = expired_bytes.saturating_add(entry.bytes);
-            false
-        }
-    });
-    let evicted = before.saturating_sub(entries.len());
-    if evicted > 0 {
-        MRF_COALESCER_COUNT.fetch_sub(evicted, Ordering::Relaxed);
-        MRF_COALESCER_BYTES.fetch_sub(expired_bytes, Ordering::Relaxed);
-        let evicted = u64::try_from(evicted).unwrap_or(u64::MAX);
-        metrics::counter!("rustfs_heal_mrf_coalescer_expired_total").increment(evicted);
-        metrics::counter!("rustfs_heal_mrf_coalescer_evictions_total").increment(evicted);
-    }
-    if entries.contains_key(&key) {
-        metrics::counter!("rustfs_heal_mrf_coalesced_total").increment(1);
-        return Err(MrfIngressResult::Coalesced);
-    }
-    let bytes = identity_estimated_bytes(&key);
-    let count_reserved = reserve(&MRF_COALESCER_COUNT, MRF_COALESCER_MAX_KEYS, 1);
-    let bytes_reserved = count_reserved && reserve(&MRF_COALESCER_BYTES, MRF_COALESCER_MAX_BYTES, bytes);
-    if !count_reserved || !bytes_reserved {
-        if count_reserved {
-            MRF_COALESCER_COUNT.fetch_sub(1, Ordering::Relaxed);
-        }
-        metrics::counter!("rustfs_heal_mrf_dropped_total", "reason" => "coalescer_full").increment(1);
-        return Err(MrfIngressResult::Dropped(MrfDropReason::CoalescerFull));
-    }
-    let lease = MrfIngressLease::new(NEXT_MRF_LEASE.fetch_add(1, Ordering::Relaxed));
-    if entries
-        .insert(
-            key,
-            IngressEntry {
-                lease,
-                expires_at: now + MRF_COALESCER_TTL,
-                bytes,
-            },
-        )
-        .is_some()
-    {
-        MRF_COALESCER_COUNT.fetch_sub(1, Ordering::Relaxed);
-        MRF_COALESCER_BYTES.fetch_sub(bytes, Ordering::Relaxed);
-        metrics::counter!("rustfs_heal_mrf_coalesced_total").increment(1);
-        return Err(MrfIngressResult::Coalesced);
-    }
-    Ok(lease)
-}
-
-fn coalescer_release(key: &MrfIdentityKey, lease: Option<MrfIngressLease>) {
-    let Some(lease) = lease else {
-        return;
-    };
-    if let Ok(mut entries) = coalescer()[key_shard(key)].lock() {
-        let should_remove = entries.get(key).is_some_and(|entry| entry.lease == lease);
-        if should_remove {
-            let bytes = entries.remove(key).map(|entry| entry.bytes).unwrap_or(0);
-            MRF_COALESCER_COUNT.fetch_sub(1, Ordering::Relaxed);
-            MRF_COALESCER_BYTES.fetch_sub(bytes, Ordering::Relaxed);
-        }
-    }
-}
-
 /// Delivery kill-switch, set from `RUSTFS_HEAL_MRF_ENABLE`. Producers check
 /// this before touching the channel so the disabled path stays allocation- and
 /// sync-free.
@@ -326,90 +122,21 @@ pub fn init_mrf_channel() -> Result<mpsc::Receiver<MrfIntent>, &'static str> {
 /// This runs on IO error paths, so it stays synchronous and cheap: one
 /// bounded allocation for the two `Arc<str>` handles plus the channel slot.
 pub fn try_send_mrf_intent(kind: MrfKind, bucket: &str, object: &str, version_id: Option<Uuid>) -> bool {
-    matches!(
-        try_send_mrf_intent_typed(kind, bucket, object, version_id, None),
-        MrfIngressResult::Enqueued
-    )
-}
-
-/// Typed ingress result. `Coalesced` means an equivalent in-flight channel
-/// intent already exists; it is not a second executable or durable admission.
-pub fn try_send_mrf_intent_typed(
-    kind: MrfKind,
-    bucket: &str,
-    object: &str,
-    version_id: Option<Uuid>,
-    scope: Option<MrfScope>,
-) -> MrfIngressResult {
    if !mrf_delivery_enabled() {
-        return MrfIngressResult::Dropped(MrfDropReason::Disabled);
+        return false;
    }
    let Some(sender) = GLOBAL_MRF_SENDER.get() else {
-        return MrfIngressResult::Dropped(MrfDropReason::Uninitialized);
-    };
-    if bucket.len() > MRF_MAX_IDENTITY_COMPONENT || object.len() > MRF_MAX_IDENTITY_COMPONENT {
-        return MrfIngressResult::Dropped(MrfDropReason::OversizedIdentity);
-    }
-    let (version_id, scope) = canonical_identity(kind, canonical_version(version_id), scope);
-    let key = MrfIdentityKey {
-        kind,
-        bucket: Arc::from(bucket),
-        object: Arc::from(object),
-        version_id,
-        scope,
-    };
-    let lease = match coalescer_admit(key.clone()) {
-        Ok(lease) => lease,
-        Err(result) => return result,
+        return false;
    };
    let intent = MrfIntent {
-        bucket: key.bucket.clone(),
-        object: key.object.clone(),
-        version_id: key.version_id,
+        bucket: Arc::from(bucket),
+        object: Arc::from(object),
+        version_id: version_id.map(|vid| *vid.as_bytes()),
        kind,
-        scope,
-        lease: Some(lease),
        enqueued_at_ms: unix_now_ms(),
        attempts: 0,
    };
-    match sender.try_send(intent) {
-        Ok(()) => MrfIngressResult::Enqueued,
-        Err(mpsc::error::TrySendError::Full(_)) => {
-            coalescer_release(&key, Some(lease));
-            metrics::counter!("rustfs_heal_mrf_dropped_total", "reason" => "channel_full").increment(1);
-            MrfIngressResult::Dropped(MrfDropReason::Full)
-        }
-        Err(mpsc::error::TrySendError::Closed(_)) => {
-            coalescer_release(&key, Some(lease));
-            MrfIngressResult::Dropped(MrfDropReason::Uninitialized)
-        }
-    }
-}
-
-/// Release the ingress key once the consumer owns the intent.
-pub fn release_mrf_intent(intent: &MrfIntent) {
-    release_mrf_identity(intent.kind, &intent.bucket, &intent.object, intent.version_id, intent.scope, intent.lease);
-}
-
-pub fn release_mrf_identity(
-    kind: MrfKind,
-    bucket: &str,
-    object: &str,
-    version_id: Option<[u8; 16]>,
-    scope: Option<MrfScope>,
-    lease: Option<MrfIngressLease>,
-) {
-    let (version_id, scope) = canonical_identity(kind, version_id, scope);
-    coalescer_release(
-        &MrfIdentityKey {
-            kind,
-            bucket: Arc::from(bucket),
-            object: Arc::from(object),
-            version_id,
-            scope,
-        },
-        lease,
-    );
+    sender.try_send(intent).is_ok()
 }

 fn unix_now_ms() -> u64 {
@@ -417,8 +144,7 @@ fn unix_now_ms() -> u64 {
    // failure would be a bug rather than something to handle here.
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
-        .ok()
-        .and_then(|d| u64::try_from(d.as_millis()).ok())
+        .map(|d| d.as_millis() as u64)
        .unwrap_or(0)
 }

@@ -489,60 +215,12 @@ mod tests {
            object: Arc::from("object"),
            version_id: Some([0u8; 16]),
            kind: MrfKind::DecodeFailure,
-            scope: None,
-            lease: None,
            enqueued_at_ms: 0,
            attempts: 0,
        };
        assert!(intent.estimated_bytes() >= intent.bucket.len() + intent.object.len());
    }

-    #[test]
-    fn ingress_duplicate_identity_coalesces_and_releases_for_retry() {
-        let key = MrfIdentityKey {
-            kind: MrfKind::DecodeFailure,
-            bucket: Arc::from("ingress-test-bucket"),
-            object: Arc::from("ingress-test-object"),
-            version_id: Some([9; 16]),
-            scope: Some(MrfScope {
-                pool_index: 3,
-                set_index: 4,
-            }),
-        };
-        let lease = coalescer_admit(key.clone()).expect("first identity should be admitted");
-        for _ in 0..999 {
-            assert_eq!(coalescer_admit(key.clone()), Err(MrfIngressResult::Coalesced));
-        }
-        coalescer_release(&key, Some(lease));
-        let retry_lease = coalescer_admit(key.clone()).expect("released identity must admit a retry");
-        coalescer_release(&key, Some(retry_lease));
-    }
-
-    #[test]
-    fn ingress_identity_preserves_kind_scope_and_version_boundaries() {
-        let (nil_version, nil_scope) = canonical_identity(
-            MrfKind::DecodeFailure,
-            Some([0; 16]),
-            Some(MrfScope {
-                pool_index: 1,
-                set_index: 2,
-            }),
-        );
-        assert_eq!(nil_version, None, "nil UUID is the unversioned identity");
-        assert!(nil_scope.is_some());
-
-        let (metadata_version, metadata_scope) = canonical_identity(
-            MrfKind::MetadataCorruption,
-            Some([7; 16]),
-            Some(MrfScope {
-                pool_index: 1,
-                set_index: 2,
-            }),
-        );
-        assert_eq!(metadata_version, None);
-        assert_eq!(metadata_scope, None);
-    }
-
    #[tokio::test]
    async fn try_send_delivers_and_respects_capacity() {
        let mut receiver = init_mrf_channel().expect("first initialization should succeed");
@@ -552,7 +230,6 @@ mod tests {
        let intent = receiver.recv().await.expect("intent should arrive");
        assert_eq!(intent.kind, MrfKind::DecodeFailure);
        assert_eq!(intent.bucket.as_ref(), "b");
-        release_mrf_intent(&intent);

        // Disable delivery: producers become no-ops.
        set_mrf_delivery_enabled(false);
@@ -562,8 +239,8 @@ mod tests {
        // Fill the bounded channel past capacity: excess intents are dropped,
        // never blocking.
        let mut accepted = 0;
-        for index in 0..(MRF_CHANNEL_CAPACITY + 64) {
-            if try_send_mrf_intent(MrfKind::PartialWrite, "b", &format!("o-{index}"), None) {
+        for _ in 0..(MRF_CHANNEL_CAPACITY + 64) {
+            if try_send_mrf_intent(MrfKind::PartialWrite, "b", "o", None) {
                accepted += 1;
            }
        }
@@ -8000,10 +8000,15 @@ impl DiskAPI for LocalDisk {
            use std::io::Write as _;

            let file_path = self.io_get_object_path(volume, path)?;
-            let lock_path = file_path.with_extension("rustfs-cas.lock");
            let path = path.to_string();
            let sync_metadata = effective_durability(volume).syncs_commit_metadata();
            return Ok(tokio::task::spawn_blocking(move || {
+                // A persistent directory lock bounds metadata growth. Removing
+                // per-target lock files can split flock ownership across inodes.
+                let lock_path = file_path
+                    .parent()
+                    .ok_or_else(|| std::io::Error::new(ErrorKind::InvalidInput, "conditional file has no parent"))?
+                    .join(".rustfs-cas.lock");
                let lock = std::fs::OpenOptions::new()
                    .create(true)
                    .truncate(false)
@@ -8068,7 +8073,25 @@ impl DiskAPI for LocalDisk {
            .map_err(DiskError::from)??);
        }

-        #[cfg(not(unix))]
+        #[cfg(windows)]
+        {
+            let file_path = self.io_get_object_path(volume, path)?;
+            let sync_metadata = effective_durability(volume).syncs_commit_metadata();
+            let publication_root = self.publication_root.clone();
+            return Ok(tokio::task::spawn_blocking(move || {
+                os::compare_and_update_control_file(
+                    &file_path,
+                    expected.as_deref(),
+                    replacement.as_deref(),
+                    sync_metadata,
+                    &publication_root,
+                )
+            })
+            .await
+            .map_err(DiskError::from)??);
+        }
+
+        #[cfg(not(any(unix, windows)))]
        {
            let _ = (volume, path, expected, replacement);
            Err(DiskError::MethodNotAllowed)
@@ -21823,9 +21846,9 @@ mod test {
        assert!(matches!(results[1].as_ref().unwrap_err(), DiskError::Io(_)));
    }

-    #[cfg(unix)]
+    #[cfg(any(unix, windows))]
    #[tokio::test]
-    async fn conditional_file_update_never_deletes_a_new_owner() {
+    async fn windows_and_unix_conditional_file_update_never_deletes_a_new_owner() {
        use tempfile::tempdir;

        let dir = tempdir().expect("temp dir should be created");
@@ -21856,8 +21879,18 @@ mod test {
            disk.read_all(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
                .await
                .expect("new owner marker should remain"),
-            owner_b
+            owner_b.clone()
        );
+        assert_eq!(
+            disk.compare_and_update_file(RUSTFS_META_BUCKET, HEALING_MARKER_PATH, Some(owner_b), None)
+                .await
+                .expect("current owner should remove marker"),
+            ConditionalFileUpdate::Updated
+        );
+        assert!(matches!(
+            disk.read_all(RUSTFS_META_BUCKET, HEALING_MARKER_PATH).await,
+            Err(DiskError::FileNotFound)
+        ));
    }

    #[cfg(unix)]
@@ -21872,7 +21905,10 @@ mod test {
        let marker_path = disk
            .get_object_path(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
            .expect("marker path should resolve");
-        let lock_path = marker_path.with_extension("rustfs-cas.lock");
+        let lock_path = marker_path
+            .parent()
+            .expect("marker path should have a parent")
+            .join(".rustfs-cas.lock");
        let lock = std::fs::OpenOptions::new()
            .create(true)
            .truncate(false)
@@ -21893,6 +21929,40 @@ mod test {
        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::WouldBlock));
    }

+    #[cfg(windows)]
+    #[tokio::test]
+    async fn windows_conditional_file_update_returns_would_block_when_marker_lock_is_contended() {
+        let dir = tempfile::tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        ensure_test_volume(&disk, RUSTFS_META_BUCKET).await;
+        let marker_path = disk
+            .get_object_path(RUSTFS_META_BUCKET, HEALING_MARKER_PATH)
+            .expect("marker path should resolve");
+        let lock_path = marker_path
+            .parent()
+            .expect("marker path should have a parent")
+            .join(".rustfs-cas.lock");
+        let lock = std::fs::OpenOptions::new()
+            .create(true)
+            .truncate(false)
+            .read(true)
+            .write(true)
+            .open(lock_path)
+            .expect("marker lock should open");
+        lock.try_lock().expect("marker lock should be held");
+
+        let err = tokio::time::timeout(
+            Duration::from_secs(1),
+            disk.compare_and_update_file(RUSTFS_META_BUCKET, HEALING_MARKER_PATH, None, Some(Bytes::from_static(b"owner"))),
+        )
+        .await
+        .expect("contended conditional update must not block")
+        .expect_err("contended conditional update must retry");
+
+        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::WouldBlock));
+    }
+
    #[cfg(target_os = "linux")]
    #[tokio::test]
    async fn replacement_io_paths_stay_under_the_mount_lease() {
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#[cfg(windows)]
+use crate::disk::ConditionalFileUpdate;
 use crate::disk::error::DiskError;
 use crate::disk::error::Result;
 use crate::disk::error_conv::to_file_error;
@@ -3458,6 +3460,89 @@ fn read_windows_relative_file(file_path: &Path, parent_guard: &ExistingBaseDirec
    Ok(Some(data))
 }

+#[cfg(windows)]
+pub(crate) fn compare_and_update_control_file(
+    file_path: &Path,
+    expected: Option<&[u8]>,
+    replacement: Option<&[u8]>,
+    sync_metadata: bool,
+    publication_root: &PublicationRoot,
+) -> io::Result<ConditionalFileUpdate> {
+    use windows_sys::{
+        Wdk::Storage::FileSystem::{
+            FILE_NON_DIRECTORY_FILE, FILE_OPEN, FILE_OPEN_IF, FILE_OPEN_REPARSE_POINT, FILE_SYNCHRONOUS_IO_NONALERT,
+        },
+        Win32::Storage::FileSystem::{
+            DELETE, FILE_ATTRIBUTE_NORMAL, FILE_READ_ATTRIBUTES, FILE_SHARE_READ, FILE_SHARE_WRITE, FILE_WRITE_DATA, SYNCHRONIZE,
+        },
+    };
+
+    let parent = file_path
+        .parent()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "conditional file has no parent"))?;
+    let parent_guard = lock_windows_directory_tree(parent, Some(parent), publication_root)?;
+    let lock = open_windows_relative(
+        parent_guard.last_handle()?,
+        std::ffi::OsStr::new(".rustfs-cas.lock"),
+        SYNCHRONIZE | FILE_READ_ATTRIBUTES | FILE_WRITE_DATA,
+        FILE_SHARE_READ | FILE_SHARE_WRITE,
+        FILE_OPEN_IF,
+        FILE_NON_DIRECTORY_FILE | FILE_OPEN_REPARSE_POINT | FILE_SYNCHRONOUS_IO_NONALERT,
+        FILE_ATTRIBUTE_NORMAL,
+        true,
+    )?;
+    validate_windows_owned_file(&lock)?;
+    match lock.as_file().try_lock() {
+        Ok(()) => {}
+        Err(std::fs::TryLockError::WouldBlock) => return Err(io::Error::from(io::ErrorKind::WouldBlock)),
+        Err(std::fs::TryLockError::Error(err)) => return Err(err),
+    }
+
+    let current = read_windows_relative_file(file_path, &parent_guard)?;
+    let matches = match (&current, expected) {
+        (None, None) => true,
+        (Some(current), Some(expected)) => current.as_slice() == expected,
+        _ => false,
+    };
+    if !matches {
+        return Ok(match current {
+            None => ConditionalFileUpdate::Missing,
+            Some(_) => ConditionalFileUpdate::Mismatch,
+        });
+    }
+
+    match replacement {
+        Some(replacement) => RenameDestinationPathGuard {
+            directory: parent.to_path_buf(),
+            _directory_guard: parent_guard,
+        }
+        .write_file_for_path_access(file_path, replacement, sync_metadata, sync_metadata)?,
+        None => {
+            let file_name = file_path
+                .file_name()
+                .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "conditional file must have a name"))?;
+            let file = open_windows_relative(
+                parent_guard.last_handle()?,
+                file_name,
+                DELETE | SYNCHRONIZE | FILE_READ_ATTRIBUTES,
+                FILE_SHARE_READ,
+                FILE_OPEN,
+                FILE_NON_DIRECTORY_FILE | FILE_OPEN_REPARSE_POINT | FILE_SYNCHRONOUS_IO_NONALERT,
+                0,
+                true,
+            )?;
+            validate_windows_owned_file(&file)?;
+            set_windows_file_delete_on_close(&file, true)?;
+            drop(file);
+            if sync_metadata {
+                fsync_dir_std(parent)?;
+            }
+        }
+    }
+
+    Ok(ConditionalFileUpdate::Updated)
+}
+
 #[cfg(windows)]
 fn open_windows_directory_component(
    parent: &WindowsDirectoryHandle,
@@ -1412,11 +1412,8 @@ pub(in crate::set_disk) async fn submit_read_repair_heal_with_submitter(

    // Reservation won: this sighting owns the repair records for the object,
    // including the durable journal intent when the caller asked for one.
-    if let Some((kind, version_uuid)) = mrf_intent
-        && let (Ok(pool_index), Ok(set_index)) = (u32::try_from(pool_index), u32::try_from(set_index))
-    {
-        let scope = rustfs_common::mrf_channel::MrfScope { pool_index, set_index };
-        let _ = rustfs_common::mrf_channel::try_send_mrf_intent_typed(kind, bucket, object, version_uuid, Some(scope));
+    if let Some((kind, version_uuid)) = mrf_intent {
+        rustfs_common::mrf_channel::try_send_mrf_intent(kind, bucket, object, version_uuid);
    }

    let mut request = rustfs_common::heal_channel::create_heal_request_with_options(
@@ -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 {
@@ -6650,23 +6637,12 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
    async fn add_partial(&self, bucket: &str, object: &str, version_id: &str) -> Result<()> {
        // MRF journal intent: partial-write recovery must survive a restart
        // (HS-01); the heal request below remains the in-memory fast path.
-        let version_uuid = if version_id.is_empty() {
-            Some(None)
-        } else {
-            uuid::Uuid::try_parse(version_id).ok().map(Some)
-        };
-        if let Some(version_uuid) = version_uuid
-            && let (Ok(pool_index), Ok(set_index)) = (u32::try_from(self.pool_index), u32::try_from(self.set_index))
-        {
-            let scope = rustfs_common::mrf_channel::MrfScope { pool_index, set_index };
-            let _ = rustfs_common::mrf_channel::try_send_mrf_intent_typed(
-                rustfs_common::mrf_channel::MrfKind::PartialWrite,
-                bucket,
-                object,
-                version_uuid,
-                Some(scope),
-            );
-        }
+        rustfs_common::mrf_channel::try_send_mrf_intent(
+            rustfs_common::mrf_channel::MrfKind::PartialWrite,
+            bucket,
+            object,
+            uuid::Uuid::try_parse(version_id).ok(),
+        );
        let mut request = rustfs_common::heal_channel::create_heal_request_with_options(
            bucket.to_string(),
            Some(object.to_string()),
@@ -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() {
@@ -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);
@@ -91,6 +91,7 @@ metrics = { workspace = true }
 base64 = { workspace = true }
 bytes = { workspace = true }
 crc-fast = { workspace = true }
+sha2 = { workspace = true }

 [dev-dependencies]
 serde_json = { workspace = true, features = ["raw_value"] }
@@ -373,6 +373,11 @@ impl ErasureSetHealer {
        set_disk_id: &str,
        buckets: &[String],
    ) -> Result<(ResumeManager, CheckpointManager)> {
+        if self.replacement_task_id.is_none() && CheckpointManager::is_blocked(&self.disk, task_id).await {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Resume task {task_id} has a blocked checkpoint"),
+            });
+        }
        // check if resume state exists
        let has_resume_state = if self.replacement_task_id.is_some() {
            ResumeManager::has_replacement_intent(&self.disk, task_id).await
@@ -119,9 +119,6 @@ struct MrfRepairNoticeTarget {
    bucket: Arc<str>,
    object: Arc<str>,
    version_id: Option<[u8; 16]>,
-    kind: rustfs_common::mrf_channel::MrfKind,
-    scope: Option<rustfs_common::mrf_channel::MrfScope>,
-    lease: Option<rustfs_common::mrf_channel::MrfIngressLease>,
 }

 #[derive(Debug, Clone)]
@@ -892,19 +889,7 @@ impl HealManager {
    }

    fn remove_mrf_repair_notice_targets_for_task(&self, task_id: &str) {
-        let targets = lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets).remove(task_id);
-        if let Some(targets) = targets {
-            for target in targets {
-                rustfs_common::mrf_channel::release_mrf_identity(
-                    target.kind,
-                    &target.bucket,
-                    &target.object,
-                    target.version_id,
-                    target.scope,
-                    target.lease,
-                );
-            }
-        }
+        lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets).remove(task_id);
    }

    fn insert_mrf_repair_notice_target(
@@ -1294,20 +1279,7 @@ impl HealManager {
        }
        self.task_aliases.lock().await.clear();
        self.retrying_heals.lock().await.clear();
-        let mrf_targets = {
-            let mut registry = lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets);
-            registry.drain().flat_map(|(_, targets)| targets).collect::<Vec<_>>()
-        };
-        for target in mrf_targets {
-            rustfs_common::mrf_channel::release_mrf_identity(
-                target.kind,
-                &target.bucket,
-                &target.object,
-                target.version_id,
-                target.scope,
-                target.lease,
-            );
-        }
+        lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets).clear();
        crate::set_heal_queue_length(0);

        // update state
@@ -1339,32 +1311,12 @@ impl HealManager {
            .await
    }

-    #[cfg(test)]
    pub(crate) async fn submit_mrf_heal_request_with_receipt(
        &self,
        request: HealRequest,
        bucket: Arc<str>,
        object: Arc<str>,
        version_id: Option<[u8; 16]>,
-    ) -> Result<HealAdmissionReceipt> {
-        let kind = match &request.heal_type {
-            HealType::Metadata { .. } => rustfs_common::mrf_channel::MrfKind::MetadataCorruption,
-            HealType::ECDecode { .. } => rustfs_common::mrf_channel::MrfKind::DecodeFailure,
-            _ => rustfs_common::mrf_channel::MrfKind::PartialWrite,
-        };
-        self.submit_mrf_heal_request_with_receipt_and_identity(request, bucket, object, version_id, kind, None, None)
-            .await
-    }
-
-    pub(crate) async fn submit_mrf_heal_request_with_receipt_and_identity(
-        &self,
-        request: HealRequest,
-        bucket: Arc<str>,
-        object: Arc<str>,
-        version_id: Option<[u8; 16]>,
-        kind: rustfs_common::mrf_channel::MrfKind,
-        scope: Option<rustfs_common::mrf_channel::MrfScope>,
-        lease: Option<rustfs_common::mrf_channel::MrfIngressLease>,
    ) -> Result<HealAdmissionReceipt> {
        self.submit_heal_request_with_receipt_alias_and_mrf_notice(
            request,
@@ -1373,9 +1325,6 @@ impl HealManager {
                bucket,
                object,
                version_id,
-                kind,
-                scope,
-                lease,
            }),
        )
        .await
@@ -1590,7 +1539,7 @@ impl HealManager {
            Self::insert_mrf_repair_notice_target(&mut targets, &task_id, target);
        }
        if let Some(displaced_task_id) = &displaced_task_id {
-            self.remove_mrf_repair_notice_targets_for_task(displaced_task_id);
+            lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets).remove(displaced_task_id);
        }
        drop(retrying_heals);
        drop(queue);
@@ -506,18 +506,7 @@ impl HealManager {
                                    &displaced_terminal,
                                )
                                .await;
-                                if let Some(targets) = lock_mrf_repair_notice_targets(&mrf_repair_notice_targets).remove(&displaced_task_id) {
-                                    for target in targets {
-                                        rustfs_common::mrf_channel::release_mrf_identity(
-                                            target.kind,
-                                            &target.bucket,
-                                            &target.object,
-                                            target.version_id,
-                                            target.scope,
-                                            target.lease,
-                                        );
-                                    }
-                                }
+                                lock_mrf_repair_notice_targets(&mrf_repair_notice_targets).remove(&displaced_task_id);
                            }
                            if matches!(admission, HealAdmissionResult::Accepted) {
                                if should_notify {
@@ -299,11 +299,7 @@ impl PriorityHealQueue {

    /// Create a deduplication key from a heal request
    pub(super) fn make_dedup_key(request: &HealRequest) -> String {
-        let base = Self::make_dedup_key_for_type(&request.heal_type);
-        match (&request.heal_type, request.options.set_key()) {
-            (HealType::Object { .. } | HealType::ECDecode { .. }, Some(scope)) => format!("{base}:scope:{scope}"),
-            _ => base,
-        }
+        Self::make_dedup_key_for_type(&request.heal_type)
    }

    pub(super) fn make_dedup_key_for_type(heal_type: &HealType) -> String {
@@ -349,8 +349,6 @@ impl HealManager {
                            let notice_targets = take_mrf_repair_notice_targets(&mrf_repair_notice_targets_clone, &task_id);
                            if successful_completion {
                                emit_mrf_repaired_events(notice_targets);
-                            } else {
-                                release_mrf_repair_notice_targets(notice_targets);
                            }
                            task_aliases_clone
                                .lock()
@@ -640,19 +638,7 @@ pub(super) fn running_heal_set_counts(active_heals: &HashMap<String, Arc<HealTas
 }

 fn remove_mrf_repair_notice_targets(registry: &Arc<StdMutex<HashMap<String, Vec<MrfRepairNoticeTarget>>>>, task_id: &str) {
-    let targets = lock_mrf_repair_notice_targets(registry).remove(task_id);
-    if let Some(targets) = targets {
-        for target in targets {
-            rustfs_common::mrf_channel::release_mrf_identity(
-                target.kind,
-                &target.bucket,
-                &target.object,
-                target.version_id,
-                target.scope,
-                target.lease,
-            );
-        }
-    }
+    lock_mrf_repair_notice_targets(registry).remove(task_id);
 }

 fn take_mrf_repair_notice_targets(
@@ -685,27 +671,6 @@ fn move_mrf_repair_notice_targets(
 fn emit_mrf_repaired_events(targets: Vec<MrfRepairNoticeTarget>) {
    for target in targets {
        rustfs_common::mrf_channel::note_mrf_repaired(&target.bucket, &target.object, target.version_id);
-        rustfs_common::mrf_channel::release_mrf_identity(
-            target.kind,
-            &target.bucket,
-            &target.object,
-            target.version_id,
-            target.scope,
-            target.lease,
-        );
-    }
-}
-
-fn release_mrf_repair_notice_targets(targets: Vec<MrfRepairNoticeTarget>) {
-    for target in targets {
-        rustfs_common::mrf_channel::release_mrf_identity(
-            target.kind,
-            &target.bucket,
-            &target.object,
-            target.version_id,
-            target.scope,
-            target.lease,
-        );
    }
 }

@@ -37,7 +37,7 @@ use crate::heal::manager::HealManager;
 use metrics::{counter, gauge};
 use rustfs_common::heal_channel::{HealAdmissionDropReason, HealAdmissionResult};
 use rustfs_common::mrf_channel::{MRF_MAX_ATTEMPTS, MrfIntent};
-use std::collections::{HashSet, VecDeque};
+use std::collections::VecDeque;
 use std::sync::Arc;
 use std::time::Duration;
 use tokio::sync::mpsc;
@@ -48,27 +48,15 @@ use crate::heal::task::{HealOptions, HealPriority, HealRequest, HealType};
 /// Journal location inside the metadata bucket, following the resume-state
 /// layout.
 pub(crate) const MRF_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal.bin";
-/// The scoped path is the authoritative snapshot for new readers and carries
-/// both v1 and v2 records. The legacy path is only a v1 compatibility mirror;
-/// older readers ignore the authoritative path, while new readers never merge
-/// the two files. This prevents a partial two-file flush from fabricating a
-/// mixed epoch.
-pub(crate) const MRF_SCOPED_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal-scoped.bin";

 /// Record format tag.
 const MRF_JOURNAL_FORMAT: u8 = 1;
 /// Record layout version.
 const MRF_JOURNAL_VERSION: u8 = 1;
-const MRF_JOURNAL_VERSION_SCOPED: u8 = 2;

 /// Fixed header size: format, version, kind, attempts, enqueued_at_ms,
 /// has_version flag.
 const MRF_RECORD_FIXED_HEAD: usize = 1 + 1 + 1 + 1 + 8 + 1;
-const MRF_MAX_IDENTITY_COMPONENT: usize = 1024;
-
-fn metric_f64(value: usize) -> f64 {
-    f64::from(u32::try_from(value).unwrap_or(u32::MAX))
-}

 #[derive(Debug, Clone)]
 pub(crate) struct MrfConsumerConfig {
@@ -113,90 +101,40 @@ impl Default for MrfConsumerConfig {
 /// incoming intent (never a resident one) and counts the loss.
 pub(crate) struct MrfQueue {
    pending: VecDeque<MrfIntent>,
-    pending_keys: HashSet<MrfQueueKey>,
    bytes: usize,
    capacity: usize,
    byte_budget: usize,
 }

-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-struct MrfQueueKey {
-    kind: rustfs_common::mrf_channel::MrfKind,
-    bucket: Arc<str>,
-    object: Arc<str>,
-    version_id: Option<[u8; 16]>,
-    scope: Option<rustfs_common::mrf_channel::MrfScope>,
-}
-
-fn queue_key(intent: &MrfIntent) -> MrfQueueKey {
-    let version_id = intent.version_id.filter(|bytes| *bytes != [0; 16]);
-    let scope = (!matches!(intent.kind, rustfs_common::mrf_channel::MrfKind::MetadataCorruption))
-        .then_some(intent.scope)
-        .flatten();
-    MrfQueueKey {
-        kind: intent.kind,
-        bucket: intent.bucket.clone(),
-        object: intent.object.clone(),
-        version_id,
-        scope,
-    }
-}
-
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
-pub(crate) enum MrfQueuePushResult {
-    Enqueued,
-    Coalesced,
-    Rejected,
-}
-
 impl MrfQueue {
    pub(crate) fn new(capacity: usize, byte_budget: usize) -> Self {
        Self {
            pending: VecDeque::new(),
-            pending_keys: HashSet::new(),
            bytes: 0,
            capacity,
            byte_budget,
        }
    }

-    pub(crate) fn try_push_typed(&mut self, intent: MrfIntent) -> MrfQueuePushResult {
-        if intent.bucket.len() > MRF_MAX_IDENTITY_COMPONENT || intent.object.len() > MRF_MAX_IDENTITY_COMPONENT {
-            counter!("rustfs_heal_mrf_dropped_total", "reason" => "identity_oversized").increment(1);
-            return MrfQueuePushResult::Rejected;
-        }
-        let key = queue_key(&intent);
-        if self.pending_keys.contains(&key) {
-            counter!("rustfs_heal_mrf_coalesced_total", "layer" => "queue").increment(1);
-            return MrfQueuePushResult::Coalesced;
-        }
+    /// Returns `false` (after counting) when either ceiling would be crossed.
+    pub(crate) fn try_push(&mut self, intent: MrfIntent) -> bool {
        let cost = intent.estimated_bytes();
        if self.pending.len() >= self.capacity || self.bytes + cost > self.byte_budget {
            counter!("rustfs_heal_mrf_dropped_total", "reason" => "queue_overflow").increment(1);
-            return MrfQueuePushResult::Rejected;
+            return false;
        }
        self.bytes += cost;
-        self.pending_keys.insert(key);
        self.pending.push_back(intent);
-        MrfQueuePushResult::Enqueued
-    }
-
-    /// Bool compatibility adapter: only a newly executable queue item is
-    /// reported as accepted; a coalesced duplicate is not durable admission.
-    #[cfg(test)]
-    pub(crate) fn try_push(&mut self, intent: MrfIntent) -> bool {
-        matches!(self.try_push_typed(intent), MrfQueuePushResult::Enqueued)
+        true
    }

    pub(crate) fn pop_front(&mut self) -> Option<MrfIntent> {
        let intent = self.pending.pop_front()?;
-        self.pending_keys.remove(&queue_key(&intent));
        self.bytes = self.bytes.saturating_sub(intent.estimated_bytes());
        Some(intent)
    }

    pub(crate) fn push_back(&mut self, intent: MrfIntent) {
-        self.pending_keys.insert(queue_key(&intent));
        self.bytes += intent.estimated_bytes();
        self.pending.push_back(intent);
    }
@@ -219,24 +157,10 @@ impl MrfQueue {
 // ---------------------------------------------------------------------------

 /// Append one encoded record to `out`.
-pub(crate) fn encode_intent(intent: &MrfIntent, out: &mut Vec<u8>) -> bool {
-    let Ok(bucket_len) = u32::try_from(intent.bucket.len()) else {
-        return false;
-    };
-    let Ok(object_len) = u32::try_from(intent.object.len()) else {
-        return false;
-    };
-    let scope = (!matches!(intent.kind, rustfs_common::mrf_channel::MrfKind::MetadataCorruption))
-        .then_some(intent.scope)
-        .flatten();
-    let version_id = intent.version_id.filter(|bytes| *bytes != [0; 16]);
+pub(crate) fn encode_intent(intent: &MrfIntent, out: &mut Vec<u8>) {
    let start = out.len();
    out.push(MRF_JOURNAL_FORMAT);
-    out.push(if scope.is_some() {
-        MRF_JOURNAL_VERSION_SCOPED
-    } else {
-        MRF_JOURNAL_VERSION
-    });
+    out.push(MRF_JOURNAL_VERSION);
    out.push(match intent.kind {
        rustfs_common::mrf_channel::MrfKind::DecodeFailure => 1,
        rustfs_common::mrf_channel::MrfKind::MetadataCorruption => 2,
@@ -244,36 +168,27 @@ pub(crate) fn encode_intent(intent: &MrfIntent, out: &mut Vec<u8>) -> bool {
    });
    out.push(intent.attempts);
    out.extend_from_slice(&intent.enqueued_at_ms.to_le_bytes());
-    match version_id {
+    match intent.version_id {
        Some(bytes) => {
            out.push(1);
            out.extend_from_slice(&bytes);
        }
        None => out.push(0),
    }
-    if let Some(scope) = scope {
-        out.extend_from_slice(&scope.pool_index.to_le_bytes());
-        out.extend_from_slice(&scope.set_index.to_le_bytes());
-    }
-    out.extend_from_slice(&bucket_len.to_le_bytes());
-    out.extend_from_slice(&object_len.to_le_bytes());
+    out.extend_from_slice(&(intent.bucket.len() as u32).to_le_bytes());
+    out.extend_from_slice(&(intent.object.len() as u32).to_le_bytes());
    out.extend_from_slice(intent.bucket.as_bytes());
    out.extend_from_slice(intent.object.as_bytes());
    let mut hasher = crc_fast::Digest::new(crc_fast::CrcAlgorithm::Crc32IsoHdlc);
    hasher.update(&out[start..]);
-    let Ok(checksum) = u32::try_from(hasher.finalize()) else {
-        out.truncate(start);
-        return false;
-    };
-    out.extend_from_slice(&checksum.to_le_bytes());
-    true
+    out.extend_from_slice(&(hasher.finalize() as u32).to_le_bytes());
 }

 fn decode_one(data: &[u8]) -> Option<(MrfIntent, usize)> {
    if data.len() < MRF_RECORD_FIXED_HEAD + 8 {
        return None;
    }
-    if data[0] != MRF_JOURNAL_FORMAT || !matches!(data[1], MRF_JOURNAL_VERSION | MRF_JOURNAL_VERSION_SCOPED) {
+    if data[0] != MRF_JOURNAL_FORMAT || data[1] != MRF_JOURNAL_VERSION {
        return None;
    }
    let kind = match data[2] {
@@ -283,38 +198,24 @@ fn decode_one(data: &[u8]) -> Option<(MrfIntent, usize)> {
        _ => return None,
    };
    let attempts = data[3];
-    let enqueued_at_ms = u64::from_le_bytes(data[4..12].try_into().ok()?);
+    let enqueued_at_ms = u64::from_le_bytes(data[4..12].try_into().expect("slice length checked"));
    let has_version = data[12] != 0;
    let mut cursor = MRF_RECORD_FIXED_HEAD;
    let version_id = if has_version {
        if data.len() < cursor + 16 {
            return None;
        }
-        let bytes: [u8; 16] = data[cursor..cursor + 16].try_into().ok()?;
+        let bytes: [u8; 16] = data[cursor..cursor + 16].try_into().expect("slice length checked");
        cursor += 16;
        Some(bytes)
    } else {
        None
    };
-    let scope = if data[1] == MRF_JOURNAL_VERSION_SCOPED {
-        if data.len() < cursor + 8 {
-            return None;
-        }
-        let pool_index = u32::from_le_bytes(data[cursor..cursor + 4].try_into().ok()?);
-        let set_index = u32::from_le_bytes(data[cursor + 4..cursor + 8].try_into().ok()?);
-        cursor += 8;
-        Some(rustfs_common::mrf_channel::MrfScope { pool_index, set_index })
-    } else {
-        None
-    };
    if data.len() < cursor + 8 {
        return None;
    }
-    let bucket_len = usize::try_from(u32::from_le_bytes(data[cursor..cursor + 4].try_into().ok()?)).ok()?;
-    let object_len = usize::try_from(u32::from_le_bytes(data[cursor + 4..cursor + 8].try_into().ok()?)).ok()?;
-    if bucket_len > MRF_MAX_IDENTITY_COMPONENT || object_len > MRF_MAX_IDENTITY_COMPONENT {
-        return None;
-    }
+    let bucket_len = u32::from_le_bytes(data[cursor..cursor + 4].try_into().expect("slice length checked")) as usize;
+    let object_len = u32::from_le_bytes(data[cursor + 4..cursor + 8].try_into().expect("slice length checked")) as usize;
    cursor += 8;
    let body_end = cursor.checked_add(bucket_len)?.checked_add(object_len)?;
    let record_end = body_end.checked_add(4)?;
@@ -323,7 +224,7 @@ fn decode_one(data: &[u8]) -> Option<(MrfIntent, usize)> {
    }
    let mut hasher = crc_fast::Digest::new(crc_fast::CrcAlgorithm::Crc32IsoHdlc);
    hasher.update(&data[..body_end]);
-    if u32::try_from(hasher.finalize()).ok()? != u32::from_le_bytes(data[body_end..record_end].try_into().ok()?) {
+    if (hasher.finalize() as u32) != u32::from_le_bytes(data[body_end..record_end].try_into().expect("slice length checked")) {
        return None;
    }
    let bucket = std::sync::Arc::from(std::str::from_utf8(&data[cursor..cursor + bucket_len]).ok()?);
@@ -334,12 +235,6 @@ fn decode_one(data: &[u8]) -> Option<(MrfIntent, usize)> {
            object,
            version_id,
            kind,
-            scope: if matches!(kind, rustfs_common::mrf_channel::MrfKind::MetadataCorruption) {
-                None
-            } else {
-                scope
-            },
-            lease: None,
            enqueued_at_ms,
            attempts,
        },
@@ -374,9 +269,9 @@ async fn journal_disks() -> Vec<DiskStore> {
    map.values().flatten().cloned().collect()
 }

-async fn read_journal(path: &str) -> Option<Vec<u8>> {
+async fn read_journal() -> Option<Vec<u8>> {
    for disk in journal_disks().await {
-        match disk.read_all(super::RUSTFS_META_BUCKET, path).await {
+        match disk.read_all(super::RUSTFS_META_BUCKET, MRF_JOURNAL_PATH).await {
            Ok(bytes) => return Some(bytes.to_vec()),
            Err(_) => continue,
        }
@@ -387,51 +282,35 @@ async fn read_journal(path: &str) -> Option<Vec<u8>> {
 /// Write the snapshot to every local disk; returns true when at least one
 /// disk accepted it, so a total write failure keeps the runtime dirty and
 /// the next tick retries the persist.
-async fn write_journal(path: &str, data: &[u8]) -> bool {
+async fn write_journal(data: &[u8]) -> bool {
    let payload = bytes::Bytes::copy_from_slice(data);
    let mut any_persisted = false;
    for disk in journal_disks().await {
-        match disk.write_all(super::RUSTFS_META_BUCKET, path, payload.clone()).await {
+        match disk
+            .write_all(super::RUSTFS_META_BUCKET, MRF_JOURNAL_PATH, payload.clone())
+            .await
+        {
            Ok(()) => any_persisted = true,
            Err(err) => warn_mrf_journal_write(&err),
        }
    }
+    if !data.is_empty() {
+        counter!("rustfs_heal_mrf_journal_fsync_total").increment(1);
+    }
+    gauge!("rustfs_heal_mrf_journal_bytes").set(data.len() as f64);
    any_persisted
 }

-async fn delete_journal(path: &str) -> bool {
-    let disks = journal_disks().await;
-    if disks.is_empty() {
-        counter!("rustfs_heal_mrf_journal_delete_failures_total").increment(1);
-        return false;
-    }
-    let mut all_deleted = true;
-    for disk in disks {
-        let result = disk
+async fn delete_journal() {
+    for disk in journal_disks().await {
+        let _ = disk
            .delete(
                super::RUSTFS_META_BUCKET,
-                path,
+                MRF_JOURNAL_PATH,
                crate::heal::storage_api::owner::EcstoreDeleteOptions::default(),
            )
            .await;
-        if let Err(err) = result {
-            // Delete is idempotent: a compatibility mirror that was never
-            // written (or was already removed) is clean, not a retry state.
-            if !matches!(err, super::DiskError::FileNotFound | super::DiskError::VolumeNotFound) {
-                all_deleted = false;
-            }
-        }
    }
-    if !all_deleted {
-        counter!("rustfs_heal_mrf_journal_delete_failures_total").increment(1);
-    }
-    all_deleted
-}
-
-async fn delete_journals() -> bool {
-    let authoritative_deleted = delete_journal(MRF_SCOPED_JOURNAL_PATH).await;
-    let legacy_deleted = delete_journal(MRF_JOURNAL_PATH).await;
-    authoritative_deleted && legacy_deleted
 }

 fn warn_mrf_journal_write(err: &super::DiskError) {
@@ -452,10 +331,7 @@ fn warn_mrf_journal_write(err: &super::DiskError) {
 pub(crate) fn build_heal_request(intent: &MrfIntent) -> HealRequest {
    let bucket = intent.bucket.to_string();
    let object = intent.object.to_string();
-    let version_id = intent
-        .version_id
-        .filter(|bytes| *bytes != [0; 16])
-        .map(|bytes| Uuid::from_bytes(bytes).to_string());
+    let version_id = intent.version_id.map(|bytes| Uuid::from_bytes(bytes).to_string());
    let (heal_type, priority) = match intent.kind {
        rustfs_common::mrf_channel::MrfKind::DecodeFailure => (
            HealType::ECDecode {
@@ -475,28 +351,18 @@ pub(crate) fn build_heal_request(intent: &MrfIntent) -> HealRequest {
            HealPriority::Normal,
        ),
    };
-    let mut options = HealOptions::default();
-    if !matches!(intent.kind, rustfs_common::mrf_channel::MrfKind::MetadataCorruption)
-        && let Some(scope) = intent.scope
-    {
-        options.pool_index = usize::try_from(scope.pool_index).ok();
-        options.set_index = usize::try_from(scope.set_index).ok();
-    }
-    let mut request = HealRequest::new(heal_type, options, priority);
+    let mut request = HealRequest::new(heal_type, HealOptions::default(), priority);
    request.source = rustfs_common::heal_channel::HealRequestSource::Mrf;
    request
 }

 async fn submit_mrf_heal_request(manager: &HealManager, intent: &MrfIntent) -> crate::Result<HealAdmissionResult> {
    let receipt = manager
-        .submit_mrf_heal_request_with_receipt_and_identity(
+        .submit_mrf_heal_request_with_receipt(
            build_heal_request(intent),
            intent.bucket.clone(),
            intent.object.clone(),
            intent.version_id,
-            intent.kind,
-            intent.scope,
-            intent.lease,
        )
        .await?;
    Ok(receipt.result)
@@ -521,38 +387,16 @@ struct MrfRuntime {
 }

 impl MrfRuntime {
-    fn snapshot(&self) -> (Vec<u8>, Vec<u8>) {
-        let mut authoritative = Vec::new();
-        let mut legacy = Vec::new();
+    fn snapshot(&self) -> Vec<u8> {
+        let mut buf = Vec::new();
        for intent in self.queue.intents() {
-            let scoped_identity =
-                !matches!(intent.kind, rustfs_common::mrf_channel::MrfKind::MetadataCorruption) && intent.scope.is_some();
-            if !encode_intent(intent, &mut authoritative) {
-                counter!("rustfs_heal_mrf_dropped_total", "reason" => "journal_identity_oversized").increment(1);
-            }
-            if !scoped_identity && !encode_intent(intent, &mut legacy) {
-                counter!("rustfs_heal_mrf_dropped_total", "reason" => "journal_identity_oversized").increment(1);
-            }
+            encode_intent(intent, &mut buf);
        }
-        (authoritative, legacy)
+        buf
    }

    async fn flush(&mut self) {
-        let (authoritative, legacy) = self.snapshot();
-        let authoritative_persisted = write_journal(MRF_SCOPED_JOURNAL_PATH, &authoritative).await;
-        if !authoritative.is_empty() {
-            counter!("rustfs_heal_mrf_journal_fsync_total").increment(1);
-        }
-        gauge!("rustfs_heal_mrf_journal_bytes").set(metric_f64(authoritative.len()));
-        // Publish the compatibility mirror only after the authoritative
-        // snapshot has reached at least one disk. This ordering prevents an
-        // old reader from observing a newer epoch that a new reader cannot
-        // see when the canonical write is unavailable.
-        let legacy_persisted = authoritative_persisted && write_journal(MRF_JOURNAL_PATH, &legacy).await;
-        // Keep dirty until both the authoritative snapshot and its
-        // compatibility mirror have been accepted; otherwise a one-sided
-        // failure would never retry the missing file.
-        let persisted = authoritative_persisted && legacy_persisted;
+        let persisted = write_journal(&self.snapshot()).await;
        self.new_since_flush = 0;
        // Keep the dirty flag when every disk write failed: a clean backlog
        // would otherwise never rewrite, losing the periodic persist retry a
@@ -560,7 +404,7 @@ impl MrfRuntime {
        if persisted {
            self.dirty = false;
        }
-        self.journal_on_disk |= authoritative_persisted || legacy_persisted;
+        self.journal_on_disk = true;
    }

    /// Drain pending intents into the heal manager until it is full, the
@@ -586,7 +430,6 @@ impl MrfRuntime {
                    intent.attempts = intent.attempts.saturating_add(1);
                    if intent.attempts >= MRF_MAX_ATTEMPTS {
                        counter!("rustfs_heal_mrf_dropped_total", "reason" => "attempts_exhausted").increment(1);
-                        rustfs_common::mrf_channel::release_mrf_intent(&intent);
                        continue;
                    }
                    self.queue.push_back(intent);
@@ -595,13 +438,11 @@ impl MrfRuntime {
                }
                Ok(HealAdmissionResult::Dropped(_)) => {
                    counter!("rustfs_heal_mrf_dropped_total", "reason" => "admission_policy").increment(1);
-                    rustfs_common::mrf_channel::release_mrf_intent(&intent);
                }
                Err(_) => {
                    intent.attempts = intent.attempts.saturating_add(1);
                    if intent.attempts >= MRF_MAX_ATTEMPTS {
                        counter!("rustfs_heal_mrf_dropped_total", "reason" => "attempts_exhausted").increment(1);
-                        rustfs_common::mrf_channel::release_mrf_intent(&intent);
                        continue;
                    }
                    self.queue.push_back(intent);
@@ -610,8 +451,8 @@ impl MrfRuntime {
                }
            }
        }
-        gauge!("rustfs_heal_mrf_queue_depth").set(metric_f64(self.queue.depth()));
-        gauge!("rustfs_heal_mrf_queue_bytes").set(metric_f64(self.queue.bytes()));
+        gauge!("rustfs_heal_mrf_queue_depth").set(self.queue.depth() as f64);
+        gauge!("rustfs_heal_mrf_queue_bytes").set(self.queue.bytes() as f64);
    }
 }

@@ -655,12 +496,7 @@ pub async fn replay_journal_once(manager: &Arc<HealManager>) -> usize {
    let config = MrfConsumerConfig::default();
    let mut queue = MrfQueue::new(config.queue_capacity, config.journal_max_bytes);
    let mut backoff_until: Option<tokio::time::Instant> = None;
-    replay_into(manager, &mut queue, &mut backoff_until).await.replayed
-}
-
-struct ReplayOutcome {
-    replayed: usize,
-    journal_on_disk: bool,
+    replay_into(manager, &mut queue, &mut backoff_until).await
 }

 /// Shared replay core: read + decode + re-arm + delete, then drain what fits.
@@ -668,25 +504,11 @@ async fn replay_into(
    manager: &Arc<HealManager>,
    queue: &mut MrfQueue,
    backoff_until: &mut Option<tokio::time::Instant>,
-) -> ReplayOutcome {
-    // The scoped file is a complete authoritative snapshot. Fall back to the
-    // legacy mirror only when the authoritative path is unavailable; merging
-    // both files could combine records from different flush epochs.
-    let data = match read_journal(MRF_SCOPED_JOURNAL_PATH).await {
-        Some(data) => data,
-        None => match read_journal(MRF_JOURNAL_PATH).await {
-            Some(data) => data,
-            None => {
-                return ReplayOutcome {
-                    replayed: 0,
-                    journal_on_disk: false,
-                };
-            }
-        },
+) -> usize {
+    let Some(data) = read_journal().await else {
+        return 0;
    };
-    let mut intents = Vec::new();
-    let (decoded, truncated) = decode_journal(&data);
-    intents.extend(decoded);
+    let (intents, truncated) = decode_journal(&data);
    if truncated > 0 {
        tracing::warn!(
            target: "rustfs::heal::mrf",
@@ -694,15 +516,12 @@ async fn replay_into(
            "MRF journal had a torn tail; truncated records were discarded"
        );
    }
-    counter!("rustfs_heal_mrf_replayed_total").increment(u64::try_from(intents.len()).unwrap_or(u64::MAX));
+    counter!("rustfs_heal_mrf_replayed_total").increment(intents.len() as u64);
    let replayed = intents.len();
    for intent in intents {
-        let result = queue.try_push_typed(intent.clone());
-        if !matches!(result, MrfQueuePushResult::Enqueued) {
-            rustfs_common::mrf_channel::release_mrf_intent(&intent);
-        }
+        queue.try_push(intent);
    }
-    let journal_on_disk = !delete_journals().await;
+    delete_journal().await;

    // Drain the replayed intents immediately; whatever the manager refuses
    // stays armed in `queue` for the consumer's retry loop.
@@ -722,10 +541,7 @@ async fn replay_into(
            }
        }
    }
-    ReplayOutcome {
-        replayed,
-        journal_on_disk,
-    }
+    replayed
 }

 /// Replay the journal, then keep draining the channel into the heal manager
@@ -743,8 +559,7 @@ async fn run_mrf_consumer(manager: Arc<HealManager>, mut receiver: mpsc::Receive

    // Replay: read the journal, re-arm intents (duplicates are merged by the
    // manager's dedup key), then drop the file so the next flush starts clean.
-    let replay = replay_into(&manager, &mut runtime.queue, &mut runtime.backoff_until).await;
-    runtime.journal_on_disk = replay.journal_on_disk;
+    replay_into(&manager, &mut runtime.queue, &mut runtime.backoff_until).await;
    // The replay deleted the journal file; anything still pending (e.g. the
    // manager was full and backoff armed) must be re-persisted by the next
    // flush or a crash before it would lose those intents.
@@ -772,14 +587,9 @@ async fn run_mrf_consumer(manager: Arc<HealManager>, mut receiver: mpsc::Receive
                    return;
                }
                for intent in batch.drain(..) {
-                    match runtime.queue.try_push_typed(intent.clone()) {
-                        MrfQueuePushResult::Enqueued => {
-                            runtime.new_since_flush += 1;
-                            runtime.dirty = true;
-                        }
-                        MrfQueuePushResult::Coalesced | MrfQueuePushResult::Rejected => {
-                            rustfs_common::mrf_channel::release_mrf_intent(&intent);
-                        }
+                    if runtime.queue.try_push(intent) {
+                        runtime.new_since_flush += 1;
+                        runtime.dirty = true;
                    }
                }
                runtime.dispatch(manager.as_ref()).await;
@@ -803,14 +613,13 @@ async fn run_mrf_consumer(manager: Arc<HealManager>, mut receiver: mpsc::Receive
                    TickAction::DeleteJournal => {
                        // All intents consumed: remove the journal so a restart
                        // replays nothing (mirrors MinIO's post-replay unlink).
-                        if delete_journals().await {
-                            runtime.journal_on_disk = false;
-                            gauge!("rustfs_heal_mrf_journal_bytes").set(0.0);
-                        }
+                        delete_journal().await;
+                        runtime.journal_on_disk = false;
+                        gauge!("rustfs_heal_mrf_journal_bytes").set(0.0);
                    }
                    TickAction::Idle => {}
                }
-                gauge!("rustfs_heal_mrf_queue_depth").set(metric_f64(runtime.queue.depth()));
+                gauge!("rustfs_heal_mrf_queue_depth").set(runtime.queue.depth() as f64);
            }
        }
    }
@@ -855,8 +664,6 @@ mod tests {
            object: StdArc::from(object),
            version_id: Some([7u8; 16]),
            kind: MrfKind::DecodeFailure,
-            scope: None,
-            lease: None,
            enqueued_at_ms: 1_700_000_000_000,
            attempts,
        }
@@ -887,101 +694,17 @@ mod tests {
    fn queue_enforces_count_and_byte_ceilings() {
        let mut queue = MrfQueue::new(2, usize::MAX);
        assert!(queue.try_push(intent("b", "o", 0)));
-        assert!(queue.try_push(intent("b", "o2", 0)));
-        assert!(!queue.try_push(intent("b", "o3", 0)), "count ceiling must drop");
+        assert!(queue.try_push(intent("b", "o", 0)));
+        assert!(!queue.try_push(intent("b", "o", 0)), "count ceiling must drop");

        let mut tiny = MrfQueue::new(usize::MAX, intent("bucket", "object", 0).estimated_bytes());
        assert!(tiny.try_push(intent("bucket", "object", 0)));
        assert!(
-            !tiny.try_push(intent("bucket", "object2", 0)),
+            !tiny.try_push(intent("bucket", "object", 0)),
            "byte budget must drop before the second intent fits"
        );
    }

-    #[test]
-    fn duplicate_mrf_intents_coalesce_to_one_execution() {
-        let mut queue = MrfQueue::new(1000, usize::MAX);
-        let mut enqueued = 0;
-        let mut coalesced = 0;
-        assert_eq!(queue.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-        enqueued += 1;
-        for _ in 0..999 {
-            match queue.try_push_typed(intent("bucket", "object", 0)) {
-                MrfQueuePushResult::Coalesced => coalesced += 1,
-                other => panic!("duplicate intent was not coalesced: {other:?}"),
-            }
-        }
-        assert_eq!(enqueued, 1);
-        assert_eq!(coalesced, 999);
-        assert_eq!(queue.depth(), 1);
-    }
-
-    #[test]
-    fn mrf_dedupe_does_not_merge_adjacent_version_pool_or_kind() {
-        let mut queue = MrfQueue::new(8, usize::MAX);
-        let mut first = intent("bucket", "object", 0);
-        first.kind = MrfKind::PartialWrite;
-        first.scope = Some(rustfs_common::mrf_channel::MrfScope {
-            pool_index: 1,
-            set_index: 1,
-        });
-        assert!(queue.try_push(first.clone()));
-        first.version_id = Some([8u8; 16]);
-        assert!(queue.try_push(first));
-        let mut other_scope = intent("bucket", "object", 0);
-        other_scope.kind = MrfKind::PartialWrite;
-        other_scope.scope = Some(rustfs_common::mrf_channel::MrfScope {
-            pool_index: 2,
-            set_index: 1,
-        });
-        assert!(queue.try_push(other_scope));
-        let mut other_kind = intent("bucket", "object", 0);
-        other_kind.kind = MrfKind::DecodeFailure;
-        other_kind.scope = None;
-        assert!(queue.try_push(other_kind));
-        assert_eq!(queue.depth(), 4);
-    }
-
-    #[test]
-    fn mrf_dedupe_full_returns_rejected_with_durable_pending() {
-        let mut queue = MrfQueue::new(1, usize::MAX);
-        assert_eq!(queue.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-        assert_eq!(queue.try_push_typed(intent("bucket", "other", 0)), MrfQueuePushResult::Rejected);
-        assert_eq!(queue.depth(), 1);
-        let mut snapshot = Vec::new();
-        assert!(encode_intent(queue.intents().next().expect("resident intent"), &mut snapshot));
-        assert!(!snapshot.is_empty(), "the resident intent remains journalable after rejection");
-    }
-
-    #[test]
-    fn mrf_dedupe_failure_releases_key_for_retry() {
-        let mut queue = MrfQueue::new(1, usize::MAX);
-        assert_eq!(queue.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-        let _failed = queue.pop_front().expect("queued intent");
-        assert_eq!(queue.try_push_typed(intent("bucket", "object", 1)), MrfQueuePushResult::Enqueued);
-        assert_eq!(queue.depth(), 1);
-    }
-
-    #[test]
-    fn mrf_dedupe_key_and_map_are_bounded() {
-        let mut queue = MrfQueue::new(2, usize::MAX);
-        assert_eq!(queue.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-        assert_eq!(queue.try_push_typed(intent("bucket", "other", 0)), MrfQueuePushResult::Enqueued);
-        assert_eq!(queue.pending_keys.len(), 2);
-        assert_eq!(queue.try_push_typed(intent("bucket", "third", 0)), MrfQueuePushResult::Rejected);
-        assert_eq!(queue.depth(), 2);
-    }
-
-    #[test]
-    fn cross_node_duplicate_execution_remains_idempotent() {
-        // Node-local ingress maps intentionally do not merge across nodes;
-        // the manager's existing identity key absorbs the duplicate later.
-        let mut node_a = MrfQueue::new(8, usize::MAX);
-        let mut node_b = MrfQueue::new(8, usize::MAX);
-        assert_eq!(node_a.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-        assert_eq!(node_b.try_push_typed(intent("bucket", "object", 0)), MrfQueuePushResult::Enqueued);
-    }
-
    #[test]
    fn journal_roundtrip_preserves_intents() {
        let intents = vec![
@@ -992,8 +715,6 @@ mod tests {
                object: StdArc::from("object/c"),
                version_id: None,
                kind: MrfKind::MetadataCorruption,
-                scope: None,
-                lease: None,
                enqueued_at_ms: 5,
                attempts: 1,
            },
@@ -1045,8 +766,6 @@ mod tests {
            object: StdArc::from("o"),
            version_id: None,
            kind: MrfKind::MetadataCorruption,
-            scope: None,
-            lease: None,
            enqueued_at_ms: 0,
            attempts: 0,
        });
@@ -1058,8 +777,6 @@ mod tests {
            object: StdArc::from("o"),
            version_id: None,
            kind: MrfKind::PartialWrite,
-            scope: None,
-            lease: None,
            enqueued_at_ms: 0,
            attempts: 0,
        });
@@ -51,6 +51,7 @@ const RESUME_STATE_FILE: &str = "ahm_resume_state.json";
 const REPLACEMENT_INTENT_FILE: &str = "ahm_replacement_intent.json";
 const RESUME_PROGRESS_FILE: &str = "ahm_progress.json";
 pub(super) const RESUME_CHECKPOINT_FILE: &str = "ahm_checkpoint.json";
+pub(super) const RESUME_CHECKPOINT_BLOCKED_FILE: &str = "ahm_checkpoint.blocked";
 const REPLACEMENT_COMPLETION_PROOF_FILE: &str = "ahm_replacement_completion_proof.json";
 const REPLACEMENT_RECOVERY_DIR: &str = "ahm-replacement";
 const REPLACEMENT_INTENT_SEAL_FILE: &str = "ahm_replacement_intent_seal";
@@ -13,26 +13,31 @@
 // limitations under the License.

 use crate::{Error, Result};
+use base64::Engine as _;
 use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
 use std::collections::HashSet;
 use std::path::Path;
 use std::sync::{Arc, Mutex};
 use std::time::{SystemTime, UNIX_EPOCH};
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex as AsyncMutex, RwLock};
 use tracing::{debug, warn};

-use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
+use super::super::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskAPI, EcstoreDiskBytes};
+use super::super::{BUCKET_META_PREFIX, DiskStore, HealDiskExt, RUSTFS_META_BUCKET};
 use super::{
-    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_FILE, delete_resume_file, path_to_str,
-    validate_resume_task_id,
+    LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, PersistThrottle, RESUME_CHECKPOINT_BLOCKED_FILE, RESUME_CHECKPOINT_FILE,
+    delete_resume_file, path_to_str, validate_resume_task_id,
 };

 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";
+const RESUME_CHECKPOINT_DIGEST_FILE: &str = "ahm_checkpoint.sha256";
+const CHECKPOINT_PER_VERSION_SCHEMA: u32 = 5;

 /// Current on-disk schema version for `ResumeCheckpoint`. Same rationale as
 /// `CURRENT_RESUME_SCHEMA`: pre-per-version dedup identities are not comparable
 /// to the new `compose_key` identities, so a stale checkpoint is discarded.
-pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
+pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 6;

 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -57,6 +62,11 @@ pub struct ResumeCheckpoint {
    pub failed_objects: HashSet<String>,
    /// skipped objects
    pub skipped_objects: HashSet<String>,
+    /// Integrity digest over the checkpoint with this field set to `None`.
+    /// Keeping it in the checkpoint makes the payload and its authentication
+    /// record one CAS generation instead of two independently-written files.
+    #[serde(default)]
+    pub integrity_digest: Option<String>,
 }

 impl ResumeCheckpoint {
@@ -70,6 +80,7 @@ impl ResumeCheckpoint {
            processed_objects: HashSet::new(),
            failed_objects: HashSet::new(),
            skipped_objects: HashSet::new(),
+            integrity_digest: None,
        }
    }

@@ -116,17 +127,111 @@ pub struct CheckpointManager {
    disk: DiskStore,
    checkpoint: Arc<RwLock<ResumeCheckpoint>>,
    throttle: Mutex<PersistThrottle>,
+    save_lock: AsyncMutex<()>,
+    last_saved: Mutex<Option<EcstoreDiskBytes>>,
 }

 impl CheckpointManager {
+    fn blocked_path(task_id: &str) -> std::path::PathBuf {
+        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}"))
+    }
+
+    /// Return whether a checkpoint was permanently isolated after a malformed
+    /// or unsupported snapshot was observed.
+    pub(crate) async fn is_blocked(disk: &DiskStore, task_id: &str) -> bool {
+        if validate_resume_task_id(task_id).is_err() {
+            return false;
+        }
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return false;
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(_) => true,
+            Err(crate::heal::DiskError::FileNotFound) => false,
+            Err(_) => true,
+        }
+    }
+
+    /// Validate the checkpoint while enumerating resumable state. This reads
+    /// the checkpoint once and also isolates malformed or unsupported data.
+    pub(crate) async fn is_resumable(disk: &DiskStore, task_id: &str) -> Result<bool> {
+        validate_resume_task_id(task_id)?;
+        if Self::is_blocked(disk, task_id).await {
+            return Err(Error::InvalidCheckpoint(format!("Resume task {task_id} has a blocked checkpoint")));
+        }
+        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
+        let Ok(path) = path_to_str(&file_path) else {
+            return Err(Error::InvalidCheckpoint("Resume checkpoint path is not valid UTF-8".to_string()));
+        };
+        match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+            Ok(bytes) if bytes.is_empty() => Ok(true),
+            Ok(bytes) => Self::load_from_data(disk.clone(), task_id, bytes.to_vec())
+                .await
+                .map(|_| true),
+            Err(crate::heal::DiskError::FileNotFound) => Ok(true),
+            Err(error) => Err(error.into()),
+        }
+    }
+
+    async fn block_invalid_snapshot(disk: &DiskStore, task_id: &str) {
+        // This marker is intentionally version-agnostic: an unsupported reader
+        // must stop selector retries until an operator cleans up the snapshot.
+        let blocked_path = Self::blocked_path(task_id);
+        let Ok(path) = path_to_str(&blocked_path) else {
+            return;
+        };
+        let result = EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path,
+            None,
+            Some(EcstoreDiskBytes::from_static(b"blocked")),
+        )
+        .await;
+        match result {
+            Ok(EcstoreConditionalFileUpdate::Updated | EcstoreConditionalFileUpdate::Mismatch) => {}
+            Ok(EcstoreConditionalFileUpdate::Missing) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = "marker target disappeared",
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+            Err(error) => warn!(
+                target: "rustfs::heal::resume",
+                event = EVENT_HEAL_CHECKPOINT_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_RESUME,
+                task_id,
+                state = "blocked_marker_write_failed",
+                error = %error,
+                "Heal checkpoint could not persist its blocked marker"
+            ),
+        }
+    }
+
    /// create new checkpoint manager
    pub async fn new(disk: DiskStore, task_id: String) -> Result<Self> {
        validate_resume_task_id(&task_id)?;
+        let checkpoint_volume = format!("{RUSTFS_META_BUCKET}/{BUCKET_META_PREFIX}");
+        if let Err(error) = EcstoreDiskAPI::make_volume(disk.as_ref(), &checkpoint_volume).await
+            && error != crate::heal::DiskError::VolumeExists
+        {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Failed to create checkpoint volume: {error}"),
+            });
+        }
        let checkpoint = ResumeCheckpoint::new(task_id);
        let manager = Self {
            disk,
            checkpoint: Arc::new(RwLock::new(checkpoint)),
            throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(None),
        };

        // save initial checkpoint
@@ -140,6 +245,7 @@ impl CheckpointManager {
                error = %e,
                "Heal checkpoint persistence failed"
            );
+            return Err(e);
        }
        Ok(manager)
    }
@@ -148,11 +254,22 @@ impl CheckpointManager {
    pub async fn load_from_disk(disk: DiskStore, task_id: &str) -> Result<Self> {
        validate_resume_task_id(task_id)?;
        let checkpoint_data = Self::read_checkpoint_file(&disk, task_id).await?;
-        let mut checkpoint: ResumeCheckpoint =
-            serde_json::from_slice(&checkpoint_data).map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to deserialize checkpoint: {e}"),
-            })?;
+        Self::load_from_data(disk, task_id, checkpoint_data).await
+    }
+
+    async fn load_from_data(disk: DiskStore, task_id: &str, checkpoint_data: Vec<u8>) -> Result<Self> {
+        validate_resume_task_id(task_id)?;
+        let mut checkpoint: ResumeCheckpoint = match serde_json::from_slice(&checkpoint_data) {
+            Ok(checkpoint) => checkpoint,
+            Err(error) => {
+                Self::block_invalid_snapshot(&disk, task_id).await;
+                return Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to deserialize checkpoint: {error}"),
+                });
+            }
+        };
        if checkpoint.task_id != task_id {
+            Self::block_invalid_snapshot(&disk, task_id).await;
            return Err(Error::TaskExecutionFailed {
                message: "Resume checkpoint task id does not match filename".to_string(),
            });
@@ -163,6 +280,7 @@ impl CheckpointManager {
        // identities. Discard the stale sets and position, then stamp the
        // current schema so the scan restarts cleanly.
        if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+            Self::block_invalid_snapshot(&disk, task_id).await;
            return Err(Error::TaskExecutionFailed {
                message: format!(
                    "Checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
@@ -170,7 +288,45 @@ impl CheckpointManager {
                ),
            });
        }
-        if checkpoint.schema_version < CURRENT_CHECKPOINT_SCHEMA {
+
+        let integrity_verified = if let Some(expected) = checkpoint.integrity_digest.as_deref() {
+            let actual = Self::checkpoint_digest(&Self::serialize_without_digest(&checkpoint)?);
+            if expected != actual {
+                Self::block_invalid_snapshot(&disk, task_id).await;
+                return Err(Error::InvalidCheckpoint(format!(
+                    "Resume checkpoint digest does not match task {task_id}"
+                )));
+            }
+            true
+        } else if checkpoint.schema_version >= CURRENT_CHECKPOINT_SCHEMA {
+            Self::block_invalid_snapshot(&disk, task_id).await;
+            return Err(Error::InvalidCheckpoint(format!(
+                "Resume checkpoint digest is missing for task {task_id}"
+            )));
+        } else {
+            let digest_path = Self::digest_path(task_id);
+            let digest_path = path_to_str(&digest_path)?;
+            match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, digest_path).await {
+                Ok(expected) => {
+                    let actual = Self::checkpoint_digest(&checkpoint_data);
+                    if expected.as_ref() != actual.as_bytes() {
+                        Self::block_invalid_snapshot(&disk, task_id).await;
+                        return Err(Error::InvalidCheckpoint(format!(
+                            "Resume checkpoint digest does not match task {task_id}"
+                        )));
+                    }
+                    true
+                }
+                Err(crate::heal::DiskError::FileNotFound) => false,
+                Err(error) => {
+                    return Err(Error::TaskExecutionFailed {
+                        message: format!("Failed to read checkpoint digest: {error}"),
+                    });
+                }
+            }
+        };
+
+        if checkpoint.schema_version < CHECKPOINT_PER_VERSION_SCHEMA || !integrity_verified {
            warn!(
                target: "rustfs::heal::resume",
                event = EVENT_HEAL_CHECKPOINT_STATE,
@@ -187,13 +343,15 @@ impl CheckpointManager {
            checkpoint.skipped_objects.clear();
            checkpoint.current_bucket_index = 0;
            checkpoint.current_object_index = 0;
-            checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
        }
+        checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;

        Ok(Self {
            disk,
            checkpoint: Arc::new(RwLock::new(checkpoint)),
            throttle: Mutex::new(PersistThrottle::new()),
+            save_lock: AsyncMutex::new(()),
+            last_saved: Mutex::new(Some(EcstoreDiskBytes::from(checkpoint_data))),
        })
    }

@@ -204,7 +362,7 @@ impl CheckpointManager {
        }
        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
        match path_to_str(&file_path) {
-            Ok(path_str) => match disk.read_all(RUSTFS_META_BUCKET, path_str).await {
+            Ok(path_str) => match HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
                Ok(data) => !data.is_empty(),
                Err(_) => false,
            },
@@ -292,6 +450,8 @@ impl CheckpointManager {

        let checkpoint_file = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));
        delete_resume_file(&self.disk, &checkpoint_file).await?;
+        delete_resume_file(&self.disk, &Self::digest_path(&task_id)).await?;
+        delete_resume_file(&self.disk, &Self::blocked_path(&task_id)).await?;

        debug!(
            target: "rustfs::heal::resume",
@@ -307,21 +467,130 @@ impl CheckpointManager {

    /// save checkpoint to disk
    async fn save_checkpoint(&self) -> Result<()> {
-        let checkpoint = self.checkpoint.read().await;
+        // Serialize saves and take the snapshot only after acquiring the lock:
+        // a slower writer must not publish a snapshot taken before a newer one.
+        let _save_guard = self.save_lock.lock().await;
+        let checkpoint = self.checkpoint.read().await.clone();
        validate_resume_task_id(&checkpoint.task_id)?;
-        let checkpoint_data = serde_json::to_vec(&*checkpoint).map_err(|e| Error::TaskExecutionFailed {
-            message: format!("Failed to serialize checkpoint: {e}"),
-        })?;
+        let unsigned_checkpoint_data = Self::serialize_without_digest(&checkpoint)?;
+        let digest = Self::checkpoint_digest(&unsigned_checkpoint_data);
+        let mut persisted_checkpoint = checkpoint.clone();
+        persisted_checkpoint.integrity_digest = Some(digest);
+        let checkpoint_data =
+            EcstoreDiskBytes::from(serde_json::to_vec(&persisted_checkpoint).map_err(|e| Error::TaskExecutionFailed {
+                message: format!("Failed to serialize checkpoint: {e}"),
+            })?);

        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{}_{}", checkpoint.task_id, RESUME_CHECKPOINT_FILE));

        let path_str = path_to_str(&file_path)?;
-        self.disk
-            .write_all(RUSTFS_META_BUCKET, path_str, checkpoint_data.into())
+        let last_saved = self
+            .last_saved
+            .lock()
+            .map_err(|_| Error::TaskExecutionFailed {
+                message: "Checkpoint save state lock is poisoned; refusing to save".to_string(),
+            })?
+            .clone();
+        let update = EcstoreDiskAPI::compare_and_update_file(
+            self.disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path_str,
+            last_saved.clone(),
+            Some(checkpoint_data.clone()),
+        )
+        .await
+        .map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to save checkpoint: {e}"),
+        })?;
+
+        let expected = match update {
+            EcstoreConditionalFileUpdate::Updated => None,
+            EcstoreConditionalFileUpdate::Missing => {
+                return Err(Error::TaskExecutionFailed {
+                    message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                });
+            }
+            EcstoreConditionalFileUpdate::Mismatch => {
+                // A healthy manager normally completes the CAS above without
+                // another read or JSON parse. Inspect only after a mismatch so
+                // corruption and future schemas cannot be overwritten blindly.
+                let existing = match HealDiskExt::read_all(self.disk.as_ref(), RUSTFS_META_BUCKET, path_str).await {
+                    Ok(existing) => existing,
+                    Err(crate::heal::DiskError::FileNotFound) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint was removed after this manager saved it; refusing to recreate it".to_string(),
+                        });
+                    }
+                    Err(error) => {
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!("Failed to inspect checkpoint after CAS mismatch: {error}"),
+                        });
+                    }
+                };
+
+                if existing.is_empty() && last_saved.is_none() {
+                    Some(existing)
+                } else {
+                    let current: ResumeCheckpoint = match serde_json::from_slice(&existing) {
+                        Ok(current) => current,
+                        Err(error) => {
+                            Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                            return Err(Error::TaskExecutionFailed {
+                                message: format!("Existing checkpoint is corrupt: {error}"),
+                            });
+                        }
+                    };
+                    if current.task_id != checkpoint.task_id {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Existing checkpoint task id does not match filename".to_string(),
+                        });
+                    }
+                    if current.schema_version > CURRENT_CHECKPOINT_SCHEMA {
+                        Self::block_invalid_snapshot(&self.disk, &checkpoint.task_id).await;
+                        return Err(Error::TaskExecutionFailed {
+                            message: format!(
+                                "Existing checkpoint schema {} is newer than supported schema {CURRENT_CHECKPOINT_SCHEMA}",
+                                current.schema_version
+                            ),
+                        });
+                    }
+                    if last_saved.as_ref().is_none_or(|saved| saved.as_ref() != existing.as_ref()) {
+                        return Err(Error::TaskExecutionFailed {
+                            message: "Checkpoint changed since this manager loaded it; refusing to overwrite newer progress"
+                                .to_string(),
+                        });
+                    }
+                    Some(existing)
+                }
+            }
+        };
+
+        if let Some(expected) = expected {
+            match EcstoreDiskAPI::compare_and_update_file(
+                self.disk.as_ref(),
+                RUSTFS_META_BUCKET,
+                path_str,
+                Some(expected),
+                Some(checkpoint_data.clone()),
+            )
            .await
            .map_err(|e| Error::TaskExecutionFailed {
-                message: format!("Failed to save checkpoint: {e}"),
-            })?;
+                message: format!("Failed to save checkpoint after CAS mismatch: {e}"),
+            })? {
+                EcstoreConditionalFileUpdate::Updated => {}
+                EcstoreConditionalFileUpdate::Missing | EcstoreConditionalFileUpdate::Mismatch => {
+                    return Err(Error::TaskExecutionFailed {
+                        message: "Checkpoint changed while saving; refusing to overwrite newer progress".to_string(),
+                    });
+                }
+            }
+        }
+
+        let mut last_saved = self.last_saved.lock().map_err(|_| Error::TaskExecutionFailed {
+            message: "Checkpoint save state lock is poisoned after save".to_string(),
+        })?;
+        *last_saved = Some(checkpoint_data);

        debug!(
            target: "rustfs::heal::resume",
@@ -341,11 +610,38 @@ impl CheckpointManager {
        let file_path = Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"));

        let path_str = path_to_str(&file_path)?;
-        disk.read_all(RUSTFS_META_BUCKET, path_str)
+        HealDiskExt::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path_str)
            .await
            .map(|bytes| bytes.to_vec())
            .map_err(|e| Error::TaskExecutionFailed {
                message: format!("Failed to read checkpoint file: {e}"),
            })
    }
+
+    fn serialize_without_digest(checkpoint: &ResumeCheckpoint) -> Result<Vec<u8>> {
+        let mut unsigned = checkpoint.clone();
+        unsigned.integrity_digest = None;
+        let mut value = serde_json::to_value(&unsigned).map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to serialize checkpoint: {e}"),
+        })?;
+        for field in ["processed_objects", "failed_objects", "skipped_objects"] {
+            let Some(values) = value.get_mut(field).and_then(serde_json::Value::as_array_mut) else {
+                return Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to canonicalize checkpoint field: {field}"),
+                });
+            };
+            values.sort_by(|left, right| left.as_str().cmp(&right.as_str()));
+        }
+        serde_json::to_vec(&value).map_err(|e| Error::TaskExecutionFailed {
+            message: format!("Failed to serialize checkpoint: {e}"),
+        })
+    }
+
+    fn checkpoint_digest(checkpoint_data: &[u8]) -> String {
+        base64::engine::general_purpose::STANDARD.encode(Sha256::digest(checkpoint_data))
+    }
+
+    fn digest_path(task_id: &str) -> std::path::PathBuf {
+        Path::new(BUCKET_META_PREFIX).join(format!("{task_id}_{RESUME_CHECKPOINT_DIGEST_FILE}"))
+    }
 }
@@ -1600,6 +1600,32 @@ async fn test_checkpoint_schema_v4_discarded_on_load() {
    temp_dir.close().expect("remove schema test directory");
 }

+#[tokio::test]
+async fn downgraded_unsigned_checkpoint_resets_untrusted_progress() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    manager.add_processed_object("victim-a".to_string()).await.unwrap();
+    manager.update_position(2, 500).await.unwrap();
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let bytes = disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path).await.unwrap();
+    let mut downgraded: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
+    downgraded["schema_version"] = serde_json::json!(CURRENT_CHECKPOINT_SCHEMA - 1);
+    downgraded.as_object_mut().unwrap().remove("integrity_digest");
+    downgraded["processed_objects"] = serde_json::json!(["victim-b"]);
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&downgraded).unwrap().into())
+        .await
+        .expect("write downgraded checkpoint");
+
+    let manager = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    let checkpoint = manager.get_checkpoint().await;
+    assert_eq!(checkpoint.schema_version, CURRENT_CHECKPOINT_SCHEMA);
+    assert_eq!(checkpoint.current_bucket_index, 0);
+    assert_eq!(checkpoint.current_object_index, 0);
+    assert!(checkpoint.processed_objects.is_empty());
+    temp_dir.close().unwrap();
+}
+
 #[tokio::test]
 async fn current_normal_resume_schema_preserves_progress() {
    let (temp_dir, disk) = schema_test_disk().await;
@@ -1675,6 +1701,369 @@ async fn future_resume_and_checkpoint_schemas_are_rejected() {
    temp_dir.close().expect("remove schema test directory");
 }

+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_non_empty_truncated_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let truncated = b"{\"schema_version\":5,\"task_id\":";
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, truncated.as_slice().into())
+        .await
+        .expect("write truncated checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a truncated checkpoint must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint is corrupt"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read truncated checkpoint fixture"),
+        truncated.as_slice()
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove checkpoint save test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_save_does_not_replace_a_future_schema_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(task_id.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+
+    let error = manager
+        .update_position(2, 7)
+        .await
+        .expect_err("a future schema must fail closed during save");
+    assert!(error.to_string().contains("Existing checkpoint schema"));
+    assert_eq!(
+        disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read future checkpoint fixture"),
+        future_bytes
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove future schema test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_digest_rejects_same_length_progress_tampering() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    manager
+        .add_processed_object("victim-a".to_string())
+        .await
+        .expect("persist checkpoint progress");
+    manager.update_position(1, 1).await.expect("flush checkpoint progress");
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let original = disk
+        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+        .await
+        .expect("read checkpoint fixture");
+    let tampered = original
+        .windows(b"victim-a".len())
+        .position(|window| window == b"victim-a")
+        .map(|index| {
+            let mut bytes = original.to_vec();
+            bytes[index..index + b"victim-a".len()].copy_from_slice(b"victim-b");
+            bytes
+        })
+        .expect("checkpoint should contain the processed object");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, tampered.into())
+        .await
+        .expect("write tampered checkpoint fixture");
+
+    assert!(CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err());
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove digest test directory");
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_survives_missing_legacy_sidecar() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    manager.update_position(2, 9).await.unwrap();
+
+    let digest_path = format!("{BUCKET_META_PREFIX}/{task_id}_ahm_checkpoint.sha256");
+    delete_resume_file(&disk, Path::new(&digest_path)).await.unwrap();
+
+    let restored = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    let checkpoint = restored.get_checkpoint().await;
+    assert_eq!(checkpoint.current_bucket_index, 2);
+    assert_eq!(checkpoint.current_object_index, 9);
+    assert!(checkpoint.integrity_digest.is_some());
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_survives_multi_object_reload() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    for index in 0..32 {
+        manager.add_processed_object(format!("processed-{index}")).await.unwrap();
+        manager.add_failed_object(format!("failed-{index}")).await.unwrap();
+        manager.add_skipped_object(format!("skipped-{index}")).await.unwrap();
+    }
+    manager.update_position(2, 9).await.unwrap();
+
+    CheckpointManager::load_from_disk(disk, &task_id)
+        .await
+        .expect("a healthy multi-object checkpoint must survive reload");
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn checkpoint_integrity_rejects_a_removed_embedded_digest() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+    manager.update_position(2, 9).await.unwrap();
+
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    let bytes = disk
+        .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+        .await
+        .expect("read checkpoint fixture");
+    let mut value: serde_json::Value = serde_json::from_slice(&bytes).unwrap();
+    value["current_object_index"] = serde_json::json!(10);
+    value.as_object_mut().unwrap().remove("integrity_digest");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, serde_json::to_vec(&value).unwrap().into())
+        .await
+        .expect("write tampered checkpoint fixture");
+
+    assert!(
+        CheckpointManager::load_from_disk(disk.clone(), &task_id).await.is_err(),
+        "a current checkpoint without its embedded digest must fail closed"
+    );
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().unwrap();
+}
+
+#[tokio::test]
+async fn new_checkpoint_manager_rebuilds_an_empty_snapshot() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &checkpoint_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty checkpoint fixture");
+
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("a new manager must rebuild an empty checkpoint");
+    manager
+        .update_position(3, 11)
+        .await
+        .expect("rebuilt checkpoint must remain writable");
+    assert!(CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty checkpoint test directory");
+}
+
+#[tokio::test]
+async fn deleted_checkpoint_is_not_recreated_by_an_old_manager() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    manager.cleanup().await.expect("delete checkpoint fixture");
+
+    let error = manager
+        .update_position(1, 2)
+        .await
+        .expect_err("an old manager must not resurrect a deleted checkpoint");
+    assert!(error.to_string().contains("removed after this manager saved it"));
+    assert!(!CheckpointManager::has_checkpoint(&disk, &task_id).await);
+    temp_dir.close().expect("remove deleted checkpoint test directory");
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn checkpoint_cleanup_leaves_no_task_specific_lock_artifact() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let lock_path = Path::new(BUCKET_META_PREFIX)
+        .join(format!("{task_id}_{RESUME_CHECKPOINT_FILE}"))
+        .with_extension("rustfs-cas.lock");
+    let lock_path = temp_dir.path().join(RUSTFS_META_BUCKET).join(lock_path);
+
+    manager.cleanup().await.expect("delete checkpoint fixture");
+
+    assert!(
+        !lock_path.exists(),
+        "successful checkpoint cleanup must not leave a task-specific lock artifact"
+    );
+}
+
+#[tokio::test]
+async fn an_empty_blocked_marker_still_blocks_resume_selection() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let manager = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create checkpoint manager");
+    let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+    disk.write_all(RUSTFS_META_BUCKET, &blocked_path, EcstoreDiskBytes::new())
+        .await
+        .expect("write empty blocked marker fixture");
+
+    assert!(CheckpointManager::is_blocked(&disk, &task_id).await);
+    assert!(CheckpointManager::is_resumable(&disk, &task_id).await.is_err());
+    // Recovery requires replacing/cleaning the snapshot, then removing the
+    // marker; ordinary selector retries are intentionally not an unlock path.
+    manager.cleanup().await.expect("clean blocked checkpoint");
+    assert!(!CheckpointManager::is_blocked(&disk, &task_id).await);
+    temp_dir.close().expect("remove empty blocked marker test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_skips_healthy_tasks_with_blocked_markers() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let tasks = [
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::new()),
+        (ResumeUtils::generate_task_id(), EcstoreDiskBytes::from_static(b"blocked")),
+    ];
+    for (task_id, marker) in &tasks {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.clone(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create healthy resume state");
+        CheckpointManager::new(disk.clone(), task_id.clone())
+            .await
+            .expect("create healthy checkpoint");
+        let checkpoint_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_FILE}");
+        let checkpoint_bytes = disk
+            .read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+            .await
+            .expect("read healthy checkpoint before blocking");
+        let marker_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        disk.write_all(RUSTFS_META_BUCKET, &marker_path, marker.clone())
+            .await
+            .expect("write blocked marker");
+
+        assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &checkpoint_path)
+                .await
+                .expect("read healthy checkpoint after blocking"),
+            checkpoint_bytes
+        );
+    }
+    temp_dir.close().expect("remove blocked selector test directory");
+}
+
+#[tokio::test]
+async fn stale_checkpoint_manager_cannot_overwrite_newer_progress() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let first = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("create first checkpoint manager");
+    let second = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load second checkpoint manager");
+
+    second
+        .update_position(4, 20)
+        .await
+        .expect("persist newer checkpoint progress");
+    let error = first
+        .update_position(1, 3)
+        .await
+        .expect_err("stale checkpoint manager must not overwrite newer progress");
+    assert!(error.to_string().contains("newer progress"));
+
+    let persisted = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("load newer checkpoint progress")
+        .get_checkpoint()
+        .await;
+    assert_eq!(persisted.current_bucket_index, 4);
+    assert_eq!(persisted.current_object_index, 20);
+    temp_dir.close().expect("remove stale manager test directory");
+}
+
+#[tokio::test]
+async fn resumable_selector_isolates_future_and_corrupt_checkpoints() {
+    let (temp_dir, disk) = schema_test_disk().await;
+    let future_task = ResumeUtils::generate_task_id();
+    let corrupt_task = ResumeUtils::generate_task_id();
+    for task_id in [&future_task, &corrupt_task] {
+        ResumeManager::new(
+            disk.clone(),
+            task_id.to_string(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            vec!["bucket".to_string()],
+        )
+        .await
+        .expect("create resumable state fixture");
+    }
+
+    let future_path = format!("{BUCKET_META_PREFIX}/{future_task}_{RESUME_CHECKPOINT_FILE}");
+    let mut future = ResumeCheckpoint::new(future_task.clone());
+    future.schema_version = CURRENT_CHECKPOINT_SCHEMA + 1;
+    let future_bytes = serde_json::to_vec(&future).expect("serialize future checkpoint fixture");
+    disk.write_all(RUSTFS_META_BUCKET, &future_path, future_bytes.clone().into())
+        .await
+        .expect("write future checkpoint fixture");
+    let corrupt_path = format!("{BUCKET_META_PREFIX}/{corrupt_task}_{RESUME_CHECKPOINT_FILE}");
+    let corrupt_bytes = b"{truncated";
+    disk.write_all(RUSTFS_META_BUCKET, &corrupt_path, corrupt_bytes.as_slice().into())
+        .await
+        .expect("write corrupt checkpoint fixture");
+
+    assert!(CheckpointManager::is_resumable(&disk, &future_task).await.is_err());
+    assert!(CheckpointManager::is_resumable(&disk, &corrupt_task).await.is_err());
+    assert!(ResumeUtils::get_resumable_tasks(&disk).await.is_err());
+    for (task_id, path, bytes) in [
+        (&future_task, future_path, future_bytes),
+        (&corrupt_task, corrupt_path, corrupt_bytes.to_vec()),
+    ] {
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &path)
+                .await
+                .expect("read isolated checkpoint bytes"),
+            bytes
+        );
+        let blocked_path = format!("{BUCKET_META_PREFIX}/{task_id}_{RESUME_CHECKPOINT_BLOCKED_FILE}");
+        assert!(
+            !disk
+                .read_all(RUSTFS_META_BUCKET, &blocked_path)
+                .await
+                .expect("read checkpoint blocked marker")
+                .is_empty()
+        );
+    }
+    temp_dir.close().expect("remove selector isolation test directory");
+}
+
 #[test]
 fn test_persist_throttle_batches_until_threshold() {
    let mut throttle = PersistThrottle::new();
@@ -21,7 +21,7 @@ use uuid::Uuid;
 use super::super::{BUCKET_META_PREFIX, DiskError, DiskStore, HealDiskExt as _, RUSTFS_META_BUCKET};
 use super::replacement::{ReplacementPhase, ReplacementRecoveryRecord};
 use super::{
-    EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
+    CheckpointManager, EVENT_HEAL_RESUME_STATE, LOG_COMPONENT_HEAL, LOG_SUBSYSTEM_RESUME, REPLACEMENT_COMPLETION_PROOF_FILE,
    REPLACEMENT_INTENT_FILE, RESUME_STATE_FILE, ResumeManager, ResumeStateFile, is_replacement_intent, path_to_str,
    replacement_recovery_corruption_for_state_load, replacement_recovery_dir, validate_resume_task_id,
 };
@@ -67,6 +67,7 @@ impl ResumeUtils {
                // Extract task ID from filename: {task_id}_ahm_resume_state.json
                if let Some(task_id) = entry.strip_suffix(&format!("_{RESUME_STATE_FILE}"))
                    && validate_resume_task_id(task_id).is_ok()
+                    && CheckpointManager::is_resumable(disk, task_id).await?
                {
                    task_ids.push(task_id.to_string());
                }
@@ -35,7 +35,6 @@ use storage_api::endpoint_index::{Endpoint, EndpointServerPools, Endpoints, Pool

 const META_BUCKET: &str = ".rustfs.sys";
 const JOURNAL_REL: &str = "buckets/.heal/mrf/journal.bin";
-const SCOPED_JOURNAL_REL: &str = "buckets/.heal/mrf/journal-scoped.bin";

 async fn heal_env() -> (Vec<std::path::PathBuf>, Arc<dyn HealStorageAPI>) {
    let env = rustfs_test_utils::TestECStoreEnv::builder()
@@ -80,18 +79,14 @@ fn journal_record(kind: u8, bucket: &str, object: &str, version: Option<[u8; 16]
    body
 }

-fn write_journal_path_to_disks(disk_paths: &[std::path::PathBuf], relative_path: &str, data: &[u8]) {
+fn write_journal_to_disks(disk_paths: &[std::path::PathBuf], data: &[u8]) {
    for path in disk_paths {
-        let journal = path.join(META_BUCKET).join(relative_path);
+        let journal = path.join(META_BUCKET).join(JOURNAL_REL);
        std::fs::create_dir_all(journal.parent().expect("journal parent")).expect("create journal dir");
        std::fs::write(&journal, data).expect("write journal fixture");
    }
 }

-fn write_journal_to_disks(disk_paths: &[std::path::PathBuf], data: &[u8]) {
-    write_journal_path_to_disks(disk_paths, JOURNAL_REL, data);
-}
-
 async fn wait_until<F, Fut>(deadline: Duration, mut probe: F) -> bool
 where
    F: FnMut() -> Fut,
@@ -192,73 +187,8 @@ async fn journal_replay_arms_intents_and_deletes_the_file() {
            .all(|path| !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()),
        "the journal file must be removed after a successful replay"
    );
-    assert!(
-        disk_paths
-            .iter()
-            .all(|path| !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()),
-        "the authoritative journal file must also be removed after replay"
-    );

    let snapshot = manager.operations_snapshot().await;
    assert_eq!(snapshot.queued_by_priority.urgent, 1, "the decode-failure record must replay as Urgent");
    assert!(snapshot.queued_by_priority.normal >= 1, "the partial-write record must replay as Normal");
 }
-
-/// A canonical snapshot and its compatibility mirror may differ after a
-/// partial flush. Replay must choose the complete canonical epoch instead of
-/// combining records that never coexisted in memory.
-#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
-#[serial]
-async fn authoritative_journal_is_not_merged_with_legacy_mirror() {
-    let (disk_paths, storage) = heal_env().await;
-    let mut endpoints: Vec<Endpoint> = disk_paths
-        .iter()
-        .map(|p| Endpoint::try_from(p.to_string_lossy().as_ref()).expect("endpoint from disk path"))
-        .collect();
-    for (i, endpoint) in endpoints.iter_mut().enumerate() {
-        endpoint.set_pool_index(0);
-        endpoint.set_set_index(0);
-        endpoint.set_disk_index(i);
-    }
-    let pool = PoolEndpoints {
-        legacy: false,
-        set_count: 1,
-        drives_per_set: endpoints.len(),
-        endpoints: Endpoints::from(endpoints),
-        cmd_line: "mrf-authoritative-test".to_string(),
-        platform: String::new(),
-    };
-    init_local_disks(EndpointServerPools::from(vec![pool]))
-        .await
-        .expect("local disks should register");
-
-    let authoritative = journal_record(1, "authoritative-bucket", "authoritative-object", None, 0);
-    let legacy = journal_record(1, "legacy-bucket", "legacy-object", None, 0);
-    write_journal_path_to_disks(&disk_paths, SCOPED_JOURNAL_REL, &authoritative);
-    write_journal_path_to_disks(&disk_paths, JOURNAL_REL, &legacy);
-
-    let manager = make_manager(storage);
-    let replayed = mrf_queue::replay_journal_once(&manager).await;
-    assert_eq!(replayed, 1, "only the authoritative snapshot epoch may replay");
-
-    let snapshot = manager.operations_snapshot().await;
-    assert_eq!(snapshot.queued_by_source.mrf, 1);
-    assert!(
-        disk_paths.iter().all(|path| {
-            !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-                && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
-        }),
-        "replay cleanup must remove both journal paths"
-    );
-
-    // A scoped-only snapshot is valid during a rollout where no legacy
-    // compatibility mirror was written. Missing legacy files must not leave
-    // the runtime in a permanent cleanup-retry state.
-    let scoped_only = journal_record(1, "scoped-only-bucket", "scoped-only-object", None, 0);
-    write_journal_path_to_disks(&disk_paths, SCOPED_JOURNAL_REL, &scoped_only);
-    assert_eq!(mrf_queue::replay_journal_once(&manager).await, 1);
-    assert!(disk_paths.iter().all(|path| {
-        !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-            && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
-    }));
-}
@@ -1375,14 +1375,13 @@ impl FolderScanner {
                            // Single-flight (backlog#1894 axis A) — the
                            // recording mode and its guarantees are pinned by
                            // corrupt_metadata_recording below.
-                            let mrf_result = rustfs_common::mrf_channel::try_send_mrf_intent_typed(
+                            let mrf_accepted = rustfs_common::mrf_channel::try_send_mrf_intent(
                                rustfs_common::mrf_channel::MrfKind::MetadataCorruption,
                                &item.bucket,
                                &object,
                                None,
-                                None,
                            );
-                            match corrupt_metadata_recording(mrf_result) {
+                            match corrupt_metadata_recording(mrf_accepted) {
                                CorruptMetadataRecording::LedgerOnly => {
                                    // Recorded as Full (retry-later): admission
                                    // for this target happens in the MRF
@@ -50,12 +50,11 @@ pub(super) enum CorruptMetadataRecording {
    ImmediateAndLedger,
 }

-pub(super) fn corrupt_metadata_recording(result: rustfs_common::mrf_channel::MrfIngressResult) -> CorruptMetadataRecording {
-    match result {
-        rustfs_common::mrf_channel::MrfIngressResult::Enqueued | rustfs_common::mrf_channel::MrfIngressResult::Coalesced => {
-            CorruptMetadataRecording::LedgerOnly
-        }
-        rustfs_common::mrf_channel::MrfIngressResult::Dropped(_) => CorruptMetadataRecording::ImmediateAndLedger,
+pub(super) fn corrupt_metadata_recording(mrf_accepted: bool) -> CorruptMetadataRecording {
+    if mrf_accepted {
+        CorruptMetadataRecording::LedgerOnly
+    } else {
+        CorruptMetadataRecording::ImmediateAndLedger
    }
 }

@@ -48,20 +48,8 @@ fn scanner_alert_wire_names_match_canonical_event_names() {
 /// the backstop survives regardless of delivery.
 #[test]
 fn corrupt_metadata_recording_maps_delivery_to_backstop() {
-    assert_eq!(
-        corrupt_metadata_recording(rustfs_common::mrf_channel::MrfIngressResult::Enqueued),
-        CorruptMetadataRecording::LedgerOnly
-    );
-    assert_eq!(
-        corrupt_metadata_recording(rustfs_common::mrf_channel::MrfIngressResult::Coalesced),
-        CorruptMetadataRecording::LedgerOnly
-    );
-    assert_eq!(
-        corrupt_metadata_recording(rustfs_common::mrf_channel::MrfIngressResult::Dropped(
-            rustfs_common::mrf_channel::MrfDropReason::Full
-        )),
-        CorruptMetadataRecording::ImmediateAndLedger
-    );
+    assert_eq!(corrupt_metadata_recording(true), CorruptMetadataRecording::LedgerOnly);
+    assert_eq!(corrupt_metadata_recording(false), CorruptMetadataRecording::ImmediateAndLedger);
 }

 fn cooldown_map_len() -> usize {
@@ -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]
@@ -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.
@@ -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"] }
@@ -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(())
 }

@@ -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) };
@@ -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);
Author	SHA1	Message	Date
houseme	5ff5a7a06e	Merge branch 'main' into cxymds/fix-1927-durable-checkpoint Signed-off-by: houseme <housemecn@gmail.com>	2026-08-23 12:27:11 +08:00
Zhengchao An	b6ba89d9e4	docs(testing): document CI gate matrix (#6412 )	2026-08-23 12:09:06 +08:00
houseme	648d5166e2	feat(allocator): replace mimalloc/libmimalloc-sys with rustfs-mimalloc/rustfs-mimalloc-sys (#6404 ) Replace the upstream xonatius/mimalloc_rust.git fork (mimalloc + libmimalloc-sys) with the published rustfs-mimalloc (v0.5.0) and rustfs-mimalloc-sys (v0.5.0) crates from crates.io. The new crates are based on mimalloc V3 (v3.5.0) and provide: - MiMalloc global allocator with safe API (collect, stats_json, process_info) - Heap management and arena operations (heap module) - Full FFI bindings to mimalloc V3 Changes: - Workspace deps: mimalloc + libmimalloc-sys (git) → rustfs-mimalloc + rustfs-mimalloc-sys (crates.io) - allocator_reclaim.rs: libmimalloc_sys::mi_collect → rustfs_mimalloc::MiMalloc::collect - memory_observability.rs: raw FFI mi_stats_get_json → MiMalloc::stats_json() - main.rs: heap ownership tests use Heap::contains() (V3 API) - deny.toml: remove xonatius/mimalloc_rust.git from allow-git Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-23 12:07:25 +08:00
houseme	84eb5aebef	fix(ecstore): remove inline write debug noise (#6408 ) * fix(ecstore): remove inline write debug noise Co-Authored-By: heihutu <heihutu@gmail.com> * fix(ecstore): satisfy warning-as-error lints Co-Authored-By: heihutu <heihutu@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-23 12:07:20 +08:00
overtrue	c58accabe7	fix(ecstore): support Windows checkpoint CAS	2026-08-23 07:58:22 +08:00
马登山	4d92835d7b	merge main and fix swift lint	2026-08-23 06:57:56 +08:00
overtrue	71db3327ad	fix(heal): reset unverified checkpoint progress	2026-08-23 06:17:31 +08:00
overtrue	68e95d497e	fix(heal): require current checkpoint digest	2026-08-23 04:19:04 +08:00
overtrue	8350d73be9	fix(storage): bound conditional file lock artifacts	2026-08-23 02:24:27 +08:00
overtrue	9a2a33a7eb	fix(heal): canonicalize checkpoint integrity digest	2026-08-23 00:09:14 +08:00
cxymds	5df26a13ba	Merge branch 'main' into cxymds/fix-1927-durable-checkpoint	2026-08-22 22:06:41 +08:00
马登山	e3a362989f	fix(heal): atomically authenticate checkpoints	2026-08-22 22:05:23 +08:00
马登山	0cd2ae20e2	fix(heal): fail closed on tampered resume checkpoints	2026-08-22 15:32:12 +08:00
cxymds	55a7fa9f03	Merge branch 'main' into cxymds/fix-1927-durable-checkpoint	2026-08-22 11:25:52 +08:00
马登山	9f51f37a0d	fix(heal): make resume checkpoints crash consistent	2026-08-21 22:04:23 +08:00