Merge branch 'main' into cxymds/fix-1936-unknown-tier

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

.github/workflows/nightly-gnu.yml

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

CLAUDE.md

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

CONTRIBUTING.md

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

Cargo.lock

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

Cargo.toml

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

crates/data-usage/src/data_usage.rs

@@ -12,8 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use serde::{Deserialize, Serialize, ser::SerializeMap as _};
+use serde::{
+    Deserialize, Serialize,
+    de::{IgnoredAny, SeqAccess, Visitor},
+    ser::SerializeMap as _,
+};
 use std::{
+    borrow::Cow,
     collections::{HashMap, HashSet},
     hash::{DefaultHasher, Hash, Hasher},
     time::{Duration, SystemTime},
@@ -48,6 +53,11 @@ pub const DATA_USAGE_OBSERVED_OBJECT_NAME: &str = ".usage.observed.json";
 // RUSTFS_COMPAT_TODO(scanner-usage-v2): keep .usage.json readable and removable during rolling upgrades from pre-v2 scanners. Remove after supported direct-upgrade sources all write .usage.v2.json.
 pub const LEGACY_DATA_USAGE_OBJECT_NAME: &str = ".usage.json";
 
+/// Fixed bucket for objects whose storage class is not in the scanner's
+/// cycle-local tier registry.  Keeping this key fixed prevents untrusted or
+/// stale tier names from growing persisted per-tier maps without bound.
+pub const UNKNOWN_TIER: &str = "UNKNOWN_TIER";
+
 /// Returns true when `existing_last_update` is ahead of `now` by more than
 /// [`USAGE_LAST_UPDATE_FUTURE_TOLERANCE`], i.e. the persisted timestamp cannot be
 /// trusted for staleness comparisons and a fresh snapshot save must be allowed.
@@ -78,12 +88,301 @@ impl TierStats {
             && self.num_objects.checked_add(u.num_objects).is_some()
     }
 
+    /// Add tier counters without allowing a counter to wrap.
+    pub fn checked_add(&self, u: &TierStats) -> Option<TierStats> {
+        Some(TierStats {
+            total_size: self.total_size.checked_add(u.total_size)?,
+            num_versions: self.num_versions.checked_add(u.num_versions)?,
+            num_objects: self.num_objects.checked_add(u.num_objects)?,
+        })
+    }
+
     /// True when this tier contributed nothing, i.e. merging it is a no-op.
     pub fn is_empty(&self) -> bool {
         self.total_size == 0 && self.num_versions == 0 && self.num_objects == 0
     }
 }
 
+/// Bounded diagnostics for objects whose tier is absent from the cycle
+/// registry. Counters are authoritative; diagnostics are only a small,
+/// redacted reconciliation aid and may be dropped at the configured caps.
+pub const UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP: usize = 64;
+pub const UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP: usize = 4096;
+pub const UNKNOWN_TIER_DIAGNOSTIC_TTL: Duration = Duration::from_secs(60 * 60);
+const UNKNOWN_TIER_DIAGNOSTIC_KEY_BYTES: usize = 256;
+const UNKNOWN_TIER_DIAGNOSTIC_INPUT_CAP: usize = UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP * 4;
+
+#[derive(Clone, Debug, Default, Serialize, PartialEq, Eq)]
+pub struct UnknownTierStats {
+    /// Logical bytes retained in the scanner's normal usage total.
+    pub unknown_bytes: u64,
+    /// Physical bytes recorded in the per-tier accounting dimension.
+    ///
+    /// Older writers only had `unknown_bytes`; decoding those snapshots keeps
+    /// this field at zero and the scanner fills it for new observations.
+    #[serde(default)]
+    pub unknown_physical_bytes: u64,
+    pub unknown_objects: u64,
+    pub unknown_versions: u64,
+    pub diagnostics_dropped: u64,
+    #[serde(default)]
+    pub diagnostics: Vec<String>,
+    #[serde(default)]
+    pub diagnostics_at: Option<SystemTime>,
+    /// A saturating update occurred; this snapshot cannot prove conservation.
+    /// The field is persisted so a restarted scanner cannot mistake a
+    /// saturated legacy aggregate for exact accounting evidence.
+    #[serde(default)]
+    pub counter_overflowed: bool,
+}
+
+#[derive(Default)]
+struct BoundedDiagnostics {
+    entries: Vec<String>,
+    dropped: u64,
+}
+
+impl<'de> Deserialize<'de> for BoundedDiagnostics {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        struct BoundedDiagnosticsVisitor;
+
+        impl<'de> Visitor<'de> for BoundedDiagnosticsVisitor {
+            type Value = BoundedDiagnostics;
+
+            fn expecting(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+                formatter.write_str("a sequence of bounded tier diagnostics")
+            }
+
+            fn visit_seq<A>(self, mut sequence: A) -> Result<Self::Value, A::Error>
+            where
+                A: SeqAccess<'de>,
+            {
+                let mut bounded = BoundedDiagnostics::default();
+                let mut bytes = 0_usize;
+                let mut inspected = 0_usize;
+                loop {
+                    if bounded.entries.len() >= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP
+                        || bytes >= UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP
+                        || inspected >= UNKNOWN_TIER_DIAGNOSTIC_INPUT_CAP
+                    {
+                        if sequence.next_element::<IgnoredAny>()?.is_none() {
+                            break;
+                        }
+                        bounded.dropped = bounded.dropped.saturating_add(1);
+                        continue;
+                    }
+                    let Some(diagnostic) = sequence.next_element::<Cow<'de, str>>()? else {
+                        break;
+                    };
+                    inspected = inspected.saturating_add(1);
+                    if !is_redacted_tier_diagnostic(&diagnostic)
+                        || bytes.saturating_add(diagnostic.len()) > UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP
+                        || bounded.entries.iter().any(|entry| entry == &diagnostic)
+                    {
+                        bounded.dropped = bounded.dropped.saturating_add(1);
+                        continue;
+                    }
+                    bytes = bytes.saturating_add(diagnostic.len());
+                    bounded.entries.push(diagnostic.into_owned());
+                }
+                Ok(bounded)
+            }
+        }
+
+        deserializer.deserialize_seq(BoundedDiagnosticsVisitor)
+    }
+}
+
+#[derive(Deserialize)]
+struct UnknownTierStatsWire {
+    #[serde(default)]
+    unknown_bytes: u64,
+    #[serde(default)]
+    unknown_physical_bytes: u64,
+    #[serde(default)]
+    unknown_objects: u64,
+    #[serde(default)]
+    unknown_versions: u64,
+    #[serde(default)]
+    diagnostics_dropped: u64,
+    #[serde(default)]
+    diagnostics: BoundedDiagnostics,
+    #[serde(default)]
+    diagnostics_at: Option<SystemTime>,
+    #[serde(default)]
+    counter_overflowed: bool,
+}
+
+fn is_redacted_tier_diagnostic(diagnostic: &str) -> bool {
+    diagnostic
+        .strip_prefix("tier-hash:")
+        .is_some_and(|digest| digest.len() == 16 && digest.bytes().all(|byte| byte.is_ascii_hexdigit()))
+}
+
+fn diagnostics_expired(at: SystemTime, now: SystemTime) -> bool {
+    now.duration_since(at).map_or(true, |age| age > UNKNOWN_TIER_DIAGNOSTIC_TTL)
+}
+
+fn checked_saturating_add(left: u64, right: u64, overflowed: &mut bool) -> u64 {
+    match left.checked_add(right) {
+        Some(value) => value,
+        None => {
+            *overflowed = true;
+            u64::MAX
+        }
+    }
+}
+
+impl<'de> Deserialize<'de> for UnknownTierStats {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let wire = UnknownTierStatsWire::deserialize(deserializer)?;
+        let mut stats = Self {
+            unknown_bytes: wire.unknown_bytes,
+            unknown_physical_bytes: wire.unknown_physical_bytes,
+            unknown_objects: wire.unknown_objects,
+            unknown_versions: wire.unknown_versions,
+            counter_overflowed: wire.counter_overflowed,
+            diagnostics_dropped: wire.diagnostics_dropped,
+            diagnostics: wire.diagnostics.entries,
+            diagnostics_at: wire.diagnostics_at,
+        };
+        stats.diagnostics_dropped =
+            checked_saturating_add(stats.diagnostics_dropped, wire.diagnostics.dropped, &mut stats.counter_overflowed);
+        if stats
+            .diagnostics_at
+            .is_some_and(|at| diagnostics_expired(at, SystemTime::now()))
+        {
+            stats.diagnostics.clear();
+            stats.diagnostics_at = None;
+        }
+        Ok(stats)
+    }
+}
+
+impl UnknownTierStats {
+    /// Record one observation where the logical and physical dimensions are
+    /// the same. Kept as a small compatibility helper for callers that only
+    /// have one size value.
+    pub fn record(&mut self, tier: &str, bytes: u64, versions: u64, objects: u64) {
+        self.record_dimensions(tier, bytes, bytes, versions, objects);
+    }
+
+    /// Record one observation without conflating logical usage with physical
+    /// tier bytes. Both counters are saturating so malformed metadata cannot
+    /// wrap an aggregate.
+    pub fn record_dimensions(&mut self, tier: &str, logical_bytes: u64, physical_bytes: u64, versions: u64, objects: u64) {
+        self.unknown_bytes = checked_saturating_add(self.unknown_bytes, logical_bytes, &mut self.counter_overflowed);
+        self.unknown_physical_bytes =
+            checked_saturating_add(self.unknown_physical_bytes, physical_bytes, &mut self.counter_overflowed);
+        self.unknown_objects = checked_saturating_add(self.unknown_objects, objects, &mut self.counter_overflowed);
+        self.unknown_versions = checked_saturating_add(self.unknown_versions, versions, &mut self.counter_overflowed);
+
+        let digest = {
+            let mut hasher = DefaultHasher::new();
+            // Bound hashing work for hostile metadata while retaining enough
+            // length/prefix entropy to reconcile repeated observations.
+            hasher.write_u64(u64::try_from(tier.len()).unwrap_or(u64::MAX));
+            let bounded = &tier.as_bytes()[..tier.len().min(UNKNOWN_TIER_DIAGNOSTIC_KEY_BYTES)];
+            hasher.write(bounded);
+            hasher.write_u8(u8::from(bounded.iter().any(|byte| byte.is_ascii_control())));
+            format!("tier-hash:{:016x}", hasher.finish())
+        };
+        let now = SystemTime::now();
+        if self.diagnostics_at.is_some_and(|at| diagnostics_expired(at, now)) {
+            self.diagnostics.clear();
+        }
+        self.diagnostics_at = Some(now);
+        if self.diagnostics.iter().any(|entry| entry == &digest) {
+            return;
+        }
+        let current_bytes: usize = self.diagnostics.iter().map(String::len).sum();
+        if self.diagnostics.len() >= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP
+            || current_bytes.saturating_add(digest.len()) > UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP
+        {
+            self.diagnostics_dropped = checked_saturating_add(1, self.diagnostics_dropped, &mut self.counter_overflowed);
+            return;
+        }
+        self.diagnostics.push(digest);
+    }
+
+    pub fn merge(&mut self, other: &Self) {
+        self.unknown_bytes = checked_saturating_add(self.unknown_bytes, other.unknown_bytes, &mut self.counter_overflowed);
+        self.unknown_physical_bytes =
+            checked_saturating_add(self.unknown_physical_bytes, other.unknown_physical_bytes, &mut self.counter_overflowed);
+        self.unknown_objects = checked_saturating_add(self.unknown_objects, other.unknown_objects, &mut self.counter_overflowed);
+        self.unknown_versions =
+            checked_saturating_add(self.unknown_versions, other.unknown_versions, &mut self.counter_overflowed);
+        self.diagnostics_dropped =
+            checked_saturating_add(self.diagnostics_dropped, other.diagnostics_dropped, &mut self.counter_overflowed);
+        self.counter_overflowed |= other.counter_overflowed;
+        let now = SystemTime::now();
+        if self.diagnostics_at.is_some_and(|at| diagnostics_expired(at, now)) {
+            self.diagnostics.clear();
+            self.diagnostics_at = None;
+        }
+        if !other.diagnostics_at.is_some_and(|at| diagnostics_expired(at, now)) {
+            for diagnostic in &other.diagnostics {
+                if !is_redacted_tier_diagnostic(diagnostic) {
+                    self.diagnostics_dropped = checked_saturating_add(1, self.diagnostics_dropped, &mut self.counter_overflowed);
+                    continue;
+                }
+                if self.diagnostics.iter().any(|entry| entry == diagnostic) {
+                    continue;
+                }
+                let current_bytes: usize = self.diagnostics.iter().map(String::len).sum();
+                if self.diagnostics.len() >= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP
+                    || current_bytes.saturating_add(diagnostic.len()) > UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP
+                {
+                    self.diagnostics_dropped = checked_saturating_add(1, self.diagnostics_dropped, &mut self.counter_overflowed);
+                    continue;
+                }
+                self.diagnostics.push(diagnostic.clone());
+            }
+        }
+        if !self.diagnostics.is_empty() {
+            self.diagnostics_at = Some(now);
+        }
+    }
+
+    pub fn fits_add(&self, other: &Self) -> bool {
+        !self.counter_overflowed
+            && !other.counter_overflowed
+            && self.unknown_bytes.checked_add(other.unknown_bytes).is_some()
+            && self
+                .unknown_physical_bytes
+                .checked_add(other.unknown_physical_bytes)
+                .is_some()
+            && self.unknown_objects.checked_add(other.unknown_objects).is_some()
+            && self.unknown_versions.checked_add(other.unknown_versions).is_some()
+            && self.diagnostics_dropped.checked_add(other.diagnostics_dropped).is_some()
+    }
+
+    pub fn checked_add(&self, other: &Self) -> Option<Self> {
+        if !self.fits_add(other) {
+            return None;
+        }
+        let mut merged = self.clone();
+        merged.merge(other);
+        Some(merged)
+    }
+
+    pub fn is_empty(&self) -> bool {
+        !self.counter_overflowed
+            && self.unknown_bytes == 0
+            && self.unknown_physical_bytes == 0
+            && self.unknown_objects == 0
+            && self.unknown_versions == 0
+            && self.diagnostics_dropped == 0
+            && self.diagnostics.is_empty()
+    }
+}
+
 #[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
 pub struct AllTierStats {
     pub tiers: HashMap<String, TierStats>,
@@ -124,6 +423,31 @@ impl AllTierStats {
             .iter()
             .all(|(tier, right)| self.tiers.get(tier).is_none_or(|left| left.fits_add(right)))
     }
+
+    /// Fold keys from an older cache that are no longer present in the
+    /// current registry into the fixed unknown bucket. Built-in storage
+    /// classes remain known even when no remote tier is configured.
+    pub fn fold_unknown_tiers<'a, I>(&mut self, known_tiers: I)
+    where
+        I: IntoIterator<Item = &'a str>,
+    {
+        let known: HashSet<&str> = known_tiers.into_iter().collect();
+        let mut unknown = self.tiers.remove(UNKNOWN_TIER).unwrap_or_default();
+        let retired_tiers: Vec<String> = self
+            .tiers
+            .keys()
+            .filter(|tier| tier.as_str() != "STANDARD" && tier.as_str() != "REDUCED_REDUNDANCY" && !known.contains(tier.as_str()))
+            .cloned()
+            .collect();
+        for tier in retired_tiers {
+            if let Some(stats) = self.tiers.remove(&tier) {
+                unknown = unknown.add(&stats);
+            }
+        }
+        if !unknown.is_empty() {
+            self.tiers.insert(UNKNOWN_TIER.to_string(), unknown);
+        }
+    }
 }
 
 /// Bucket target usage info provides replication statistics
@@ -211,6 +535,10 @@ pub struct DataUsageInfo {
     /// tier exists, so an absent value means "not accounted", never "zero".
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub tier_stats: Option<AllTierStats>,
+    /// Bounded diagnostics and separate logical/physical counters for objects
+    /// classified into [`UNKNOWN_TIER`].
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub unknown_tier_stats: Option<UnknownTierStats>,
 
     /// Total number of buckets in this cluster
     pub buckets_count: u64,
@@ -307,6 +635,48 @@ pub struct DiskUsageStatus {
     pub snapshot_exists: bool,
 }
 
+/// Independent conservation evidence for a scanner summary.
+///
+/// The totals are maintained while objects are accounted and are deliberately
+/// not reconstructed from the tier map at publish time. `*_known` records the
+/// portion represented by the corresponding accounting dimension.
+#[derive(Debug, Default, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
+pub struct TierAccountingProof {
+    pub logical_total: u64,
+    pub logical_known: u64,
+    pub physical_total: u64,
+    pub physical_known: u64,
+    #[serde(default)]
+    pub overflowed: bool,
+}
+
+impl TierAccountingProof {
+    pub fn checked_add(self, other: Self) -> Option<Self> {
+        if self.overflowed || other.overflowed {
+            return None;
+        }
+        Some(Self {
+            logical_total: self.logical_total.checked_add(other.logical_total)?,
+            logical_known: self.logical_known.checked_add(other.logical_known)?,
+            physical_total: self.physical_total.checked_add(other.physical_total)?,
+            physical_known: self.physical_known.checked_add(other.physical_known)?,
+            overflowed: false,
+        })
+    }
+
+    pub fn saturating_add(&mut self, other: Self) {
+        let Some(merged) = (*self).checked_add(other) else {
+            self.logical_total = self.logical_total.saturating_add(other.logical_total);
+            self.logical_known = self.logical_known.saturating_add(other.logical_known);
+            self.physical_total = self.physical_total.saturating_add(other.physical_total);
+            self.physical_known = self.physical_known.saturating_add(other.physical_known);
+            self.overflowed = true;
+            return;
+        };
+        *self = merged;
+    }
+}
+
 /// Size summary for a single object or group of objects
 #[derive(Debug, Default, Clone)]
 pub struct SizeSummary {
@@ -336,6 +706,10 @@ pub struct SizeSummary {
     pub repl_target_stats: HashMap<String, ReplTargetSizeSummary>,
     /// Per-tier accounting, keyed by storage class or remote tier name
     pub tier_stats: HashMap<String, TierStats>,
+    /// Counters and bounded diagnostics for unknown tiers in this summary.
+    pub unknown_tier_stats: UnknownTierStats,
+    /// Independent logical/physical conservation evidence.
+    pub tier_accounting_proof: TierAccountingProof,
 }
 
 /// Replication target size summary
@@ -427,6 +801,10 @@ impl<'de> Deserialize<'de> for SizeHistogram {
 }
 
 impl SizeHistogram {
+    pub fn is_empty(&self) -> bool {
+        self.0.iter().all(|value| *value == 0)
+    }
+
     pub fn add(&mut self, size: u64) {
         let intervals = [
             (0, 1024 - 1),                              // LESS_THAN_1024_B
@@ -541,6 +919,10 @@ impl<'de> Deserialize<'de> for VersionsHistogram {
 }
 
 impl VersionsHistogram {
+    pub fn is_empty(&self) -> bool {
+        self.0.iter().all(|value| *value == 0)
+    }
+
     pub fn add(&mut self, count: u64) {
         let intervals = [
             (0, 0),            // UNVERSIONED
@@ -681,6 +1063,13 @@ pub struct DataUsageEntry {
     /// observed tier-classified objects.
     #[serde(default)]
     pub all_tier_stats: Option<AllTierStats>,
+    /// Bounded unknown-tier reconciliation state for this cache entry.
+    #[serde(default)]
+    pub unknown_tier_stats: Option<UnknownTierStats>,
+    /// Optional conservation proof. Missing values are legacy/unproven, not
+    /// zero-valued evidence.
+    #[serde(default)]
+    pub tier_accounting_proof: Option<TierAccountingProof>,
 }
 
 impl Serialize for DataUsageEntry {
@@ -691,7 +1080,9 @@ impl Serialize for DataUsageEntry {
         // Keep entries map-encoded so older readers can ignore fields appended
         // by newer scanner versions during rolling upgrades. The derived
         // (array) encoding made any appended field a decode error for them.
-        let mut state = serializer.serialize_map(Some(11))?;
+        let mut state = serializer.serialize_map(Some(
+            11 + usize::from(self.unknown_tier_stats.is_some()) + usize::from(self.tier_accounting_proof.is_some()),
+        ))?;
         state.serialize_entry("children", &self.children)?;
         state.serialize_entry("size", &self.size)?;
         state.serialize_entry("objects", &self.objects)?;
@@ -703,11 +1094,35 @@ impl Serialize for DataUsageEntry {
         state.serialize_entry("compacted", &self.compacted)?;
         state.serialize_entry("failed_objects", &self.failed_objects)?;
         state.serialize_entry("all_tier_stats", &self.all_tier_stats)?;
+        // Keep the legacy no-unknown shape byte-for-byte stable. Once unknown
+        // accounting exists, append its map field before the optional proof.
+        if let Some(unknown_tier_stats) = self.unknown_tier_stats.as_ref() {
+            state.serialize_entry("unknown_tier_stats", unknown_tier_stats)?;
+        }
+        if let Some(proof) = self.tier_accounting_proof.as_ref() {
+            state.serialize_entry("tier_accounting_proof", proof)?;
+        }
         state.end()
     }
 }
 
 impl DataUsageEntry {
+    fn has_local_usage(&self) -> bool {
+        self.size != 0
+            || self.objects != 0
+            || self.versions != 0
+            || self.delete_markers != 0
+            || self.failed_objects != 0
+            || self.obj_sizes.0.iter().any(|value| *value != 0)
+            || self.obj_versions.0.iter().any(|value| *value != 0)
+            || self.replication_stats.as_ref().is_some_and(|stats| !stats.is_empty())
+            || self
+                .all_tier_stats
+                .as_ref()
+                .is_some_and(|stats| stats.tiers.values().any(|tier| !tier.is_empty()))
+            || self.unknown_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty())
+    }
+
     pub fn add_child(&mut self, hash: &DataUsageHash) {
         if self.children.contains(&hash.key()) {
             return;
@@ -716,6 +1131,8 @@ impl DataUsageEntry {
     }
 
     pub fn merge(&mut self, other: &DataUsageEntry) {
+        let self_had_local_usage = self.has_local_usage();
+        let other_had_local_usage = other.has_local_usage();
         self.objects += other.objects;
         self.versions += other.versions;
         self.delete_markers += other.delete_markers;
@@ -744,6 +1161,29 @@ impl DataUsageEntry {
         if let Some(o_tiers) = other.all_tier_stats.as_ref().filter(|tiers| !tiers.is_empty()) {
             self.all_tier_stats.get_or_insert_with(AllTierStats::new).merge(o_tiers);
         }
+        if let Some(other_unknown) = other.unknown_tier_stats.as_ref() {
+            self.unknown_tier_stats
+                .get_or_insert_with(UnknownTierStats::default)
+                .merge(other_unknown);
+        }
+
+        self.tier_accounting_proof = match (self.tier_accounting_proof, other.tier_accounting_proof) {
+            (Some(mut left), Some(right)) => {
+                left.saturating_add(right);
+                Some(left)
+            }
+            (None, Some(right)) if !self_had_local_usage => Some(right),
+            (Some(left), None) if !other_had_local_usage => Some(left),
+            (None, None) => None,
+            _ => None,
+        };
+        // A saturated unknown-tier counter invalidates conservation evidence;
+        // never let an otherwise valid proof hide that loss of precision.
+        if self.unknown_tier_stats.as_ref().is_some_and(|stats| stats.counter_overflowed)
+            && let Some(proof) = self.tier_accounting_proof.as_mut()
+        {
+            proof.overflowed = true;
+        }
 
         self.obj_sizes.merge_from(&other.obj_sizes);
         self.obj_versions.merge_from(&other.obj_versions);
@@ -757,6 +1197,15 @@ impl DataUsageEntry {
         self.all_tier_stats.get_or_insert_with(AllTierStats::new).add_sizes(tiers);
     }
 
+    pub fn add_unknown_tier_stats(&mut self, stats: &UnknownTierStats) {
+        if stats.is_empty() {
+            return;
+        }
+        self.unknown_tier_stats
+            .get_or_insert_with(UnknownTierStats::default)
+            .merge(stats);
+    }
+
     pub fn checked_merge(&mut self, other: &DataUsageEntry) -> bool {
         let scalar_counts_fit = self.objects.checked_add(other.objects).is_some()
             && self.versions.checked_add(other.versions).is_some()
@@ -820,8 +1269,21 @@ impl DataUsageEntry {
             (_, None) | (None, Some(_)) => true,
             (Some(left), Some(right)) => left.fits_merge(right),
         };
+        let unknown_tier_stats_fit = match (&self.unknown_tier_stats, &other.unknown_tier_stats) {
+            (None, None) => true,
+            (Some(left), None) => !left.counter_overflowed,
+            (None, Some(right)) => !right.counter_overflowed,
+            (Some(left), Some(right)) => left.fits_add(right),
+        };
 
-        if !scalar_counts_fit || !histograms_fit || !replication_fits || !tier_stats_fit {
+        let proof_fit = match (self.tier_accounting_proof, other.tier_accounting_proof) {
+            (Some(left), Some(right)) => left.checked_add(right).is_some(),
+            (Some(proof), None) | (None, Some(proof)) => !proof.overflowed,
+            (None, None) => true,
+        };
+
+        if !scalar_counts_fit || !histograms_fit || !replication_fits || !tier_stats_fit || !unknown_tier_stats_fit || !proof_fit
+        {
             return false;
         }
         self.merge(other);
@@ -1339,6 +1801,7 @@ impl DataUsageCache {
             delete_markers_total_count: flat.delete_markers as u64,
             objects_total_size: flat.size as u64,
             tier_stats: flat.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+            unknown_tier_stats: flat.unknown_tier_stats.filter(|stats| !stats.is_empty()),
             buckets_count: u64::try_from(buckets.len()).unwrap_or(u64::MAX),
             buckets_usage,
             usage_snapshot_complete: self.info.snapshot_complete,
@@ -1766,6 +2229,20 @@ impl SizeSummary {
         self.replica_count = self.replica_count.saturating_add(other.replica_count);
         self.pending_count = self.pending_count.saturating_add(other.pending_count);
         self.failed_count = self.failed_count.saturating_add(other.failed_count);
+        self.unknown_tier_stats.merge(&other.unknown_tier_stats);
+        self.tier_accounting_proof.saturating_add(other.tier_accounting_proof);
+        if self.unknown_tier_stats.counter_overflowed {
+            self.tier_accounting_proof.overflowed = true;
+        }
+
+        // A disk/bucket aggregate is assembled from many object summaries.
+        // Keep the per-tier dimension in lockstep with the scalar counters;
+        // dropping this map here would recreate the original silent-loss bug
+        // at the cross-disk merge boundary.
+        for (tier, stats) in &other.tier_stats {
+            let entry = self.tier_stats.entry(tier.clone()).or_default();
+            *entry = entry.add(stats);
+        }
 
         // Merge replication target stats
         for (target, stats) in &other.repl_target_stats {
@@ -1878,6 +2355,59 @@ mod tests {
         );
     }
 
+    #[test]
+    fn retired_tier_stats_fold_into_fixed_unknown_bucket() {
+        let mut stats = AllTierStats::default();
+        stats.tiers.insert(
+            "RETIRED".to_string(),
+            TierStats {
+                total_size: 9,
+                num_versions: 2,
+                num_objects: 1,
+            },
+        );
+        stats.tiers.insert(
+            "WARM".to_string(),
+            TierStats {
+                total_size: 4,
+                num_versions: 1,
+                num_objects: 1,
+            },
+        );
+
+        stats.fold_unknown_tiers(["WARM"]);
+
+        assert!(!stats.tiers.contains_key("RETIRED"));
+        assert_eq!(stats.tiers.get("WARM").map(|v| v.total_size), Some(4));
+        assert_eq!(stats.tiers.get(UNKNOWN_TIER).map(|v| v.total_size), Some(9));
+    }
+
+    #[test]
+    fn unknown_tier_stats_deserialization_keeps_diagnostics_bounded() {
+        #[derive(Serialize)]
+        struct RawUnknownTierStats {
+            unknown_bytes: u64,
+            diagnostics: Vec<String>,
+            diagnostics_dropped: u64,
+        }
+
+        let mut diagnostics = vec!["tier-hash:0123456789abcdef".to_string(); UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP + 8];
+        diagnostics.push("raw-tier-name-that-must-not-be-exposed".to_string());
+        diagnostics.push("x".repeat(UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP + 1));
+        let encoded = rmp_serde::to_vec_named(&RawUnknownTierStats {
+            unknown_bytes: 7,
+            diagnostics,
+            diagnostics_dropped: 3,
+        })
+        .expect("unknown tier stats should encode");
+        let decoded: UnknownTierStats = rmp_serde::from_slice(&encoded).expect("unknown tier stats should decode");
+
+        assert_eq!(decoded.unknown_bytes, 7);
+        assert_eq!(decoded.diagnostics.len(), 1);
+        assert!(decoded.diagnostics_dropped >= 3);
+        assert!(decoded.diagnostics.iter().all(|entry| is_redacted_tier_diagnostic(entry)));
+    }
+
     #[test]
     fn checked_merge_rejects_overflowing_tier_totals() {
         let mut left = tier_entry(
@@ -1901,6 +2431,57 @@ mod tests {
         assert_eq!(left.all_tier_stats.expect("left is untouched").tiers["WARM"].total_size, u64::MAX);
     }
 
+    #[test]
+    fn checked_merge_rejects_unknown_counter_overflow_on_either_side() {
+        let overflowed = DataUsageEntry {
+            unknown_tier_stats: Some(UnknownTierStats {
+                counter_overflowed: true,
+                ..Default::default()
+            }),
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: 1,
+                logical_known: 1,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let mut left = DataUsageEntry::default();
+        assert!(!left.checked_merge(&overflowed));
+
+        let mut left = overflowed.clone();
+        assert!(!left.checked_merge(&DataUsageEntry::default()));
+    }
+
+    #[test]
+    fn checked_merge_rejects_one_sided_overflowed_proof_without_mutation() {
+        let mut left = DataUsageEntry {
+            size: 1,
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: 1,
+                logical_known: 1,
+                overflowed: true,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let legacy = DataUsageEntry {
+            size: 2,
+            ..Default::default()
+        };
+
+        assert!(!left.checked_merge(&legacy));
+        assert_eq!(left.size, 1);
+        assert_eq!(
+            left.tier_accounting_proof,
+            Some(TierAccountingProof {
+                logical_total: 1,
+                logical_known: 1,
+                overflowed: true,
+                ..Default::default()
+            })
+        );
+    }
+
     /// Entry shape released before per-tier accounting, using the derived
     /// (array) encoding those writers produced.
     #[derive(Serialize, Deserialize)]
@@ -1940,6 +2521,72 @@ mod tests {
         assert_eq!(legacy.objects, 0);
     }
 
+    #[test]
+    fn tier_accounting_proof_is_optional_and_map_encoded() {
+        let entry = DataUsageEntry {
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: 11,
+                logical_known: 11,
+                physical_total: 7,
+                physical_known: 7,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let encoded = rmp_serde::to_vec(&entry).expect("proof-bearing entry should encode");
+        let decoded: DataUsageEntry = rmp_serde::from_slice(&encoded).expect("proof-bearing entry should decode");
+        assert_eq!(decoded.tier_accounting_proof, entry.tier_accounting_proof);
+
+        let legacy = LegacyEntry {
+            children: DataUsageHashMap::default(),
+            size: 12,
+            objects: 3,
+            versions: 4,
+            delete_markers: 1,
+            obj_sizes: SizeHistogram::default(),
+            obj_versions: VersionsHistogram::default(),
+            replication_stats: None,
+            compacted: false,
+            failed_objects: 2,
+        };
+        let legacy_bytes = rmp_serde::to_vec(&legacy).expect("legacy entry should encode");
+        let decoded: DataUsageEntry = rmp_serde::from_slice(&legacy_bytes).expect("legacy entry should decode");
+        assert!(decoded.tier_accounting_proof.is_none());
+    }
+
+    #[test]
+    fn empty_entry_merge_preserves_accounting_proof() {
+        let mut entry = DataUsageEntry {
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: 7,
+                logical_known: 7,
+                physical_total: 7,
+                physical_known: 7,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        entry.merge(&DataUsageEntry::default());
+        assert!(entry.tier_accounting_proof.is_some());
+
+        let mut empty = DataUsageEntry::default();
+        empty.merge(&entry);
+        assert_eq!(empty.tier_accounting_proof, entry.tier_accounting_proof);
+    }
+
+    #[test]
+    fn unknown_counter_overflow_is_not_empty_and_survives_roundtrip() {
+        let stats = UnknownTierStats {
+            counter_overflowed: true,
+            ..Default::default()
+        };
+        assert!(!stats.is_empty());
+        let encoded = rmp_serde::to_vec_named(&stats).expect("overflow marker should encode");
+        let decoded: UnknownTierStats = rmp_serde::from_slice(&encoded).expect("overflow marker should decode");
+        assert!(decoded.counter_overflowed);
+        assert!(!decoded.is_empty());
+    }
+
     #[test]
     fn legacy_array_encoded_entries_still_load() {
         let legacy = LegacyEntry {
@@ -2708,6 +3355,32 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_dui_filters_empty_unknown_tier_usage_from_the_flattened_tree() {
+        let root_hash = hash_path("root");
+        let bucket_hash = hash_path("bucket-a");
+        let mut cache = DataUsageCache {
+            info: DataUsageCacheInfo {
+                name: "root".to_string(),
+                ..Default::default()
+            },
+            ..Default::default()
+        };
+        cache.replace_hashed(&root_hash, &None, &DataUsageEntry::default());
+
+        let child = DataUsageEntry {
+            objects: 1,
+            unknown_tier_stats: Some(UnknownTierStats::default()),
+            ..Default::default()
+        };
+        cache.replace_hashed(&bucket_hash, &Some(root_hash), &child);
+
+        let info = cache.dui("root", &["bucket-a".to_string()]);
+
+        assert_eq!(info.objects_total_count, 1);
+        assert!(info.unknown_tier_stats.is_none());
+    }
+
     #[test]
     fn test_data_usage_entry_merge_preserves_replication_targets() {
         let mut base = DataUsageEntry {

crates/ecstore/src/api/mod.rs

@@ -456,11 +456,13 @@ pub mod rpc {
         ScannerBucketListing, ScannerPeerActivity, TONIC_RPC_PREFIX, TonicInterceptor, build_put_file_auth_trailer,
         check_and_record_signed_rpc_nonce, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
         gen_tonic_signature_interceptor, node_service_time_out_client, node_service_time_out_client_no_auth,
-        normalize_tonic_rpc_audience, set_tonic_canonical_body_digest, sign_ns_scanner_capability, sign_put_file_capability,
-        sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers,
-        tonic_rpc_auth_failure_reason, verify_put_file_auth_trailer, verify_put_file_capability, verify_rpc_signature,
-        verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest, verify_tonic_mutation_body_digest,
-        verify_tonic_rpc_response_proof, verify_tonic_rpc_signature, verify_tonic_rpc_signature_with_bootstrap,
+        normalize_tonic_rpc_audience, set_tonic_canonical_body_digest, sign_ns_scanner_capability,
+        sign_ns_scanner_capability_with_tier_registry_generation, sign_put_file_capability, sign_tonic_rpc_response_proof,
+        tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers, tonic_rpc_auth_failure_reason,
+        verify_ns_scanner_capability, verify_ns_scanner_capability_with_tier_registry_generation, verify_put_file_auth_trailer,
+        verify_put_file_capability, verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
+        verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
+        verify_tonic_rpc_signature_with_bootstrap,
     };
 }
 

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

@@ -94,6 +94,7 @@ const REPLAY_CACHE_AUTO_MEMORY_PERCENT: u64 = 13;
 const REPLAY_CACHE_AUTO_RPC_RPS_PER_CPU: usize = 4096;
 const REPLAY_CACHE_AUTO_MAX_CAPACITY: usize = 33_554_432;
 const NS_SCANNER_CAPABILITY_AUTH_DOMAIN: &[u8] = b"rustfs-ns-scanner-capability-v3";
+const NS_SCANNER_TIER_REGISTRY_GENERATION_AUTH_DOMAIN: &[u8] = b"rustfs-ns-scanner-tier-registry-generation-v1";
 pub const TONIC_RPC_PREFIX: &str = "/node_service.NodeService";
 static INTERNODE_RPC_SIGNATURE_STRICT: LazyLock<bool> = LazyLock::new(|| {
     get_env_bool(
@@ -636,40 +637,79 @@ pub fn verify_put_file_capability(challenge: Uuid, server_epoch: Uuid, version:
         .map_err(|_| std::io::Error::new(std::io::ErrorKind::PermissionDenied, "Invalid put_file capability proof"))
 }
 
-fn update_ns_scanner_capability_mac(mac: &mut HmacSha256, challenge: Uuid, server_epoch: Uuid) {
+fn update_ns_scanner_capability_mac(
+    mac: &mut HmacSha256,
+    challenge: Uuid,
+    server_epoch: Uuid,
+    supports_tier_registry_generation: bool,
+) {
     mac.update(NS_SCANNER_CAPABILITY_AUTH_DOMAIN);
     mac.update(&NS_SCANNER_PROTOCOL_VERSION.to_be_bytes());
     mac.update(challenge.as_bytes());
     mac.update(server_epoch.as_bytes());
+    if supports_tier_registry_generation {
+        // The optional response capability is part of the authenticated
+        // scope. A proxy cannot turn an old/unsupported peer into a worker
+        // that receives generation-fenced scanner work.
+        mac.update(NS_SCANNER_TIER_REGISTRY_GENERATION_AUTH_DOMAIN);
+    }
 }
 
-fn generate_ns_scanner_capability_proof(secret: &str, challenge: Uuid, server_epoch: Uuid) -> std::io::Result<Vec<u8>> {
+fn generate_ns_scanner_capability_proof(
+    secret: &str,
+    challenge: Uuid,
+    server_epoch: Uuid,
+    supports_tier_registry_generation: bool,
+) -> std::io::Result<Vec<u8>> {
     if challenge.is_nil() || server_epoch.is_nil() {
         return Err(std::io::Error::other("Invalid namespace scanner capability scope"));
     }
     let mut mac =
         <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
-    update_ns_scanner_capability_mac(&mut mac, challenge, server_epoch);
+    update_ns_scanner_capability_mac(&mut mac, challenge, server_epoch, supports_tier_registry_generation);
     Ok(mac.finalize().into_bytes().to_vec())
 }
 
-fn verify_ns_scanner_capability_proof(secret: &str, challenge: Uuid, server_epoch: Uuid, proof: &[u8]) -> std::io::Result<()> {
+fn verify_ns_scanner_capability_proof(
+    secret: &str,
+    challenge: Uuid,
+    server_epoch: Uuid,
+    proof: &[u8],
+    supports_tier_registry_generation: bool,
+) -> std::io::Result<()> {
     if challenge.is_nil() || server_epoch.is_nil() {
         return Err(std::io::Error::other("Invalid namespace scanner capability scope"));
     }
     let mut mac =
         <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
-    update_ns_scanner_capability_mac(&mut mac, challenge, server_epoch);
+    update_ns_scanner_capability_mac(&mut mac, challenge, server_epoch, supports_tier_registry_generation);
     mac.verify_slice(proof)
         .map_err(|_| std::io::Error::new(std::io::ErrorKind::PermissionDenied, "Invalid namespace scanner capability proof"))
 }
 
 pub fn sign_ns_scanner_capability(challenge: Uuid, server_epoch: Uuid) -> std::io::Result<Vec<u8>> {
-    generate_ns_scanner_capability_proof(&get_shared_secret()?, challenge, server_epoch)
+    sign_ns_scanner_capability_with_tier_registry_generation(challenge, server_epoch, false)
 }
 
 pub fn verify_ns_scanner_capability(challenge: Uuid, server_epoch: Uuid, proof: &[u8]) -> std::io::Result<()> {
-    verify_ns_scanner_capability_proof(&get_shared_secret()?, challenge, server_epoch, proof)
+    verify_ns_scanner_capability_with_tier_registry_generation(challenge, server_epoch, proof, false)
+}
+
+pub fn sign_ns_scanner_capability_with_tier_registry_generation(
+    challenge: Uuid,
+    server_epoch: Uuid,
+    supports_tier_registry_generation: bool,
+) -> std::io::Result<Vec<u8>> {
+    generate_ns_scanner_capability_proof(&get_shared_secret()?, challenge, server_epoch, supports_tier_registry_generation)
+}
+
+pub fn verify_ns_scanner_capability_with_tier_registry_generation(
+    challenge: Uuid,
+    server_epoch: Uuid,
+    proof: &[u8],
+    supports_tier_registry_generation: bool,
+) -> std::io::Result<()> {
+    verify_ns_scanner_capability_proof(&get_shared_secret()?, challenge, server_epoch, proof, supports_tier_registry_generation)
 }
 
 #[derive(Clone, Copy)]
@@ -1709,13 +1749,28 @@ mod tests {
         let secret = "test-scanner-capability-secret";
         let challenge = Uuid::new_v4();
         let server_epoch = Uuid::new_v4();
-        let proof =
-            generate_ns_scanner_capability_proof(secret, challenge, server_epoch).expect("capability proof should be generated");
+        let proof = generate_ns_scanner_capability_proof(secret, challenge, server_epoch, false)
+            .expect("capability proof should be generated");
 
-        assert!(verify_ns_scanner_capability_proof(secret, challenge, server_epoch, &proof).is_ok());
-        assert!(verify_ns_scanner_capability_proof(secret, Uuid::new_v4(), server_epoch, &proof).is_err());
-        assert!(verify_ns_scanner_capability_proof(secret, challenge, Uuid::new_v4(), &proof).is_err());
-        assert!(verify_ns_scanner_capability_proof("different-secret", challenge, server_epoch, &proof).is_err());
+        assert!(verify_ns_scanner_capability_proof(secret, challenge, server_epoch, &proof, false).is_ok());
+        assert!(verify_ns_scanner_capability_proof(secret, Uuid::new_v4(), server_epoch, &proof, false).is_err());
+        assert!(verify_ns_scanner_capability_proof(secret, challenge, Uuid::new_v4(), &proof, false).is_err());
+        assert!(verify_ns_scanner_capability_proof("different-secret", challenge, server_epoch, &proof, false).is_err());
+    }
+
+    #[test]
+    fn namespace_scanner_capability_proof_binds_tier_registry_generation_support() {
+        let secret = "test-scanner-capability-secret";
+        let challenge = Uuid::new_v4();
+        let server_epoch = Uuid::new_v4();
+        let proof = generate_ns_scanner_capability_proof(secret, challenge, server_epoch, true)
+            .expect("generation capability proof should be generated");
+
+        assert!(verify_ns_scanner_capability_proof(secret, challenge, server_epoch, &proof, true).is_ok());
+        assert!(verify_ns_scanner_capability_proof(secret, challenge, server_epoch, &proof, false).is_err());
+        let legacy = generate_ns_scanner_capability_proof(secret, challenge, server_epoch, false)
+            .expect("legacy capability proof should be generated");
+        assert!(verify_ns_scanner_capability_proof(secret, challenge, server_epoch, &legacy, true).is_err());
     }
 
     /// Security regression for GHSA-r5qv-rc46-hv8q (internode RPC fail-closed,

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

@@ -13,17 +13,18 @@
 // limitations under the License.
 
 use crate::cluster::rpc::{
-    build_auth_headers, build_put_file_auth_trailer, verify_ns_scanner_capability, verify_put_file_capability,
+    build_auth_headers, build_put_file_auth_trailer, verify_ns_scanner_capability_with_tier_registry_generation,
+    verify_put_file_capability,
 };
 use crate::disk::error::{Error, Result};
 use crate::disk::{FileReader, FileWriter};
 use crate::storage_api_contracts::internode::{
     NS_SCANNER_BODY_SHA256_QUERY, NS_SCANNER_CAPABILITY_CHALLENGE_QUERY, NS_SCANNER_CYCLE_QUERY, NS_SCANNER_LEADER_EPOCH_QUERY,
     NS_SCANNER_PROTOCOL_VERSION, NS_SCANNER_PROTOCOL_VERSION_QUERY, NS_SCANNER_REQUEST_ID_QUERY, NS_SCANNER_SERVER_EPOCH_QUERY,
-    NS_SCANNER_SESSION_ID_QUERY, NS_SCANNER_SESSION_SEQUENCE_QUERY, NsScannerCapabilityResponse, PUT_FILE_AUTH_QUERY,
-    PUT_FILE_AUTH_V1, PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY, PUT_FILE_CAPABILITY_VERSION,
-    PUT_FILE_NONCE_QUERY, PUT_FILE_SERVER_EPOCH_QUERY, PutFileCapabilityResponse, WALK_DIR_BODY_SHA256_QUERY,
-    WALK_DIR_STREAM_COMPLETION_QUERY, WALK_DIR_STREAM_COMPLETION_V1,
+    NS_SCANNER_SESSION_ID_QUERY, NS_SCANNER_SESSION_SEQUENCE_QUERY, NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY,
+    NsScannerCapabilityResponse, PUT_FILE_AUTH_QUERY, PUT_FILE_AUTH_V1, PUT_FILE_CAPABILITY_CHALLENGE_QUERY,
+    PUT_FILE_CAPABILITY_QUERY, PUT_FILE_CAPABILITY_VERSION, PUT_FILE_NONCE_QUERY, PUT_FILE_SERVER_EPOCH_QUERY,
+    PutFileCapabilityResponse, WALK_DIR_BODY_SHA256_QUERY, WALK_DIR_STREAM_COMPLETION_QUERY, WALK_DIR_STREAM_COMPLETION_V1,
 };
 use async_trait::async_trait;
 use http::{HeaderMap, HeaderValue, Method, header::CONTENT_TYPE};
@@ -137,6 +138,12 @@ fn put_file_capability_status_is_legacy(status: u16) -> bool {
     status == 404
 }
 
+fn ns_scanner_capability_error_allows_legacy(error: &Error) -> bool {
+    [400, 404, 405, 426]
+        .into_iter()
+        .any(|status| error.is_internode_http_status(status))
+}
+
 #[derive(Debug, Clone, Copy, Eq, PartialEq)]
 #[allow(
     dead_code,
@@ -220,6 +227,7 @@ pub struct NsScannerStreamRequest {
 #[derive(Debug, Clone)]
 pub struct NsScannerCapabilityRequest {
     pub endpoint: String,
+    pub supports_tier_registry_generation: bool,
 }
 
 /// Data-plane stream opener used by `RemoteDisk`.
@@ -252,6 +260,15 @@ pub trait InternodeDataTransport: Send + Sync + std::fmt::Debug {
     async fn probe_ns_scanner(&self, _request: NsScannerCapabilityRequest) -> Result<Uuid> {
         Err(Error::MethodNotAllowed)
     }
+    async fn probe_ns_scanner_capability(&self, request: NsScannerCapabilityRequest) -> Result<NsScannerCapabilityResponse> {
+        let server_epoch = self.probe_ns_scanner(request).await?;
+        Ok(NsScannerCapabilityResponse {
+            version: NS_SCANNER_PROTOCOL_VERSION,
+            server_epoch,
+            proof: Vec::new(),
+            supports_tier_registry_generation: None,
+        })
+    }
     // Interface facet nobody calls yet: every transport implements both, but no
     // caller negotiates on them. Kept for the internode transport split
     // (backlog#1350); deleting them would delete the seam and six impls.
@@ -335,27 +352,44 @@ impl InternodeDataTransport for TcpHttpInternodeDataTransport {
     }
 
     async fn probe_ns_scanner(&self, request: NsScannerCapabilityRequest) -> Result<Uuid> {
-        let challenge = Uuid::new_v4();
-        let url = build_ns_scanner_capability_url(&request, challenge);
-        let mut headers = msgpack_headers();
-        build_auth_headers(&url, &Method::GET, &mut headers)?;
-        let reader = HttpReader::new(url, Method::GET, headers, None).await?;
-        let mut body = Vec::new();
-        reader
-            .take(u64::try_from(NS_SCANNER_MAX_CAPABILITY_RESPONSE_SIZE + 1).unwrap_or(u64::MAX))
-            .read_to_end(&mut body)
-            .await?;
-        if body.is_empty() || body.len() > NS_SCANNER_MAX_CAPABILITY_RESPONSE_SIZE {
-            return Err(Error::other("invalid remote namespace scanner capability response size"));
+        Ok(self.probe_ns_scanner_capability(request).await?.server_epoch)
+    }
+
+    async fn probe_ns_scanner_capability(&self, request: NsScannerCapabilityRequest) -> Result<NsScannerCapabilityResponse> {
+        if request.supports_tier_registry_generation {
+            return match self.probe_ns_scanner_capability_once(&request).await {
+                Ok(response) => Ok(response),
+                Err(marked_error) if ns_scanner_capability_error_allows_legacy(&marked_error) => {
+                    // A v3 peer may reject the additive query marker, ignore
+                    // it, or return its legacy proof. Retry once without the
+                    // marker and only downgrade after that legacy response is
+                    // authenticated; an unverified epoch is never trusted.
+                    let legacy_request = NsScannerCapabilityRequest {
+                        endpoint: request.endpoint.clone(),
+                        supports_tier_registry_generation: false,
+                    };
+                    match self.probe_ns_scanner_capability_once(&legacy_request).await {
+                        Ok(mut response) => {
+                            response.supports_tier_registry_generation = None;
+                            Ok(response)
+                        }
+                        Err(legacy_error) if ns_scanner_capability_error_allows_legacy(&legacy_error) => {
+                            // Some old deployments expose only the legacy
+                            // protocol response (or advertise 426). Treat
+                            // the pair as an explicit unsupported result so
+                            // the scanner can use its coordinator fallback.
+                            Err(Error::MethodNotAllowed)
+                        }
+                        Err(_) => Err(marked_error),
+                    }
+                }
+                // A server failure, network failure, or authentication error
+                // is not evidence of an old parser. Do not issue an
+                // unauthenticated legacy probe or silently downgrade.
+                Err(marked_error) => Err(marked_error),
+            };
         }
-        let response: NsScannerCapabilityResponse =
-            rmp_serde::from_slice(&body).map_err(|_| Error::other("invalid remote namespace scanner capability response"))?;
-        if response.version != NS_SCANNER_PROTOCOL_VERSION || response.server_epoch.is_nil() {
-            return Err(Error::other("incompatible remote namespace scanner capability response"));
-        }
-        verify_ns_scanner_capability(challenge, response.server_epoch, &response.proof)
-            .map_err(|err| Error::other(format!("remote namespace scanner capability authentication failed: {err}")))?;
-        Ok(response.server_epoch)
+        self.probe_ns_scanner_capability_once(&request).await
     }
 
     fn name(&self) -> &'static str {
@@ -368,6 +402,53 @@ impl InternodeDataTransport for TcpHttpInternodeDataTransport {
 }
 
 impl TcpHttpInternodeDataTransport {
+    async fn probe_ns_scanner_capability_once(
+        &self,
+        request: &NsScannerCapabilityRequest,
+    ) -> Result<NsScannerCapabilityResponse> {
+        let challenge = Uuid::new_v4();
+        let url = build_ns_scanner_capability_url(request, challenge);
+        let mut headers = msgpack_headers();
+        build_auth_headers(&url, &Method::GET, &mut headers)?;
+        let reader = HttpReader::new(url, Method::GET, headers, None).await?;
+        let mut body = Vec::new();
+        reader
+            .take(u64::try_from(NS_SCANNER_MAX_CAPABILITY_RESPONSE_SIZE + 1).unwrap_or(u64::MAX))
+            .read_to_end(&mut body)
+            .await?;
+        if body.is_empty() || body.len() > NS_SCANNER_MAX_CAPABILITY_RESPONSE_SIZE {
+            return Err(Error::other("invalid remote namespace scanner capability response size"));
+        }
+        let mut response: NsScannerCapabilityResponse =
+            rmp_serde::from_slice(&body).map_err(|_| Error::other("invalid remote namespace scanner capability response"))?;
+        if response.version != NS_SCANNER_PROTOCOL_VERSION || response.server_epoch.is_nil() {
+            return Err(Error::other("incompatible remote namespace scanner capability response"));
+        }
+        if let Err(err) = verify_ns_scanner_capability_with_tier_registry_generation(
+            challenge,
+            response.server_epoch,
+            &response.proof,
+            request.supports_tier_registry_generation,
+        ) {
+            // A permissive older peer can ignore the additive marker and
+            // return a valid legacy-scope proof with HTTP 200. Accept that
+            // response only after independently authenticating the legacy
+            // scope; all other verification failures remain fail-closed.
+            if request.supports_tier_registry_generation && ns_scanner_capability_legacy_proof_is_valid(challenge, &response) {
+                response.supports_tier_registry_generation = None;
+                return Ok(response);
+            }
+            return Err(Error::other(format!("remote namespace scanner capability authentication failed: {err}")));
+        }
+        // The proof authenticates the requested capability scope, not the
+        // optional response field. Derive the client-facing bit from that
+        // verified scope so an intermediary cannot strip or rewrite the field
+        // and force a silent downgrade after a successful generation-bound
+        // handshake.
+        normalize_ns_scanner_capability_response(&mut response, request.supports_tier_registry_generation);
+        Ok(response)
+    }
+
     async fn put_file_auth_capability(&self, endpoint: &str) -> Result<Option<Uuid>> {
         resolve_put_file_auth_capability(endpoint, || async {
             tokio::time::timeout(PUT_FILE_CAPABILITY_PROBE_TIMEOUT, self.probe_put_file_auth(endpoint))
@@ -649,6 +730,14 @@ fn build_walk_dir_url(request: &WalkDirStreamRequest) -> String {
     )
 }
 
+fn normalize_ns_scanner_capability_response(response: &mut NsScannerCapabilityResponse, requested_generation_support: bool) {
+    response.supports_tier_registry_generation = requested_generation_support.then_some(true);
+}
+
+fn ns_scanner_capability_legacy_proof_is_valid(challenge: Uuid, response: &NsScannerCapabilityResponse) -> bool {
+    verify_ns_scanner_capability_with_tier_registry_generation(challenge, response.server_epoch, &response.proof, false).is_ok()
+}
+
 fn build_ns_scanner_url(request: &NsScannerStreamRequest) -> String {
     let body_sha256 = hex_simd::encode_to_string(Sha256::digest(&request.body), hex_simd::AsciiCase::Lower);
     format!(
@@ -675,13 +764,18 @@ fn build_ns_scanner_url(request: &NsScannerStreamRequest) -> String {
 
 fn build_ns_scanner_capability_url(request: &NsScannerCapabilityRequest, challenge: Uuid) -> String {
     format!(
-        "{}{}?{}={}&{}={}",
+        "{}{}?{}={}&{}={}{}",
         request.endpoint,
         NS_SCANNER_PATH,
         NS_SCANNER_PROTOCOL_VERSION_QUERY,
         NS_SCANNER_PROTOCOL_VERSION,
         NS_SCANNER_CAPABILITY_CHALLENGE_QUERY,
-        challenge
+        challenge,
+        if request.supports_tier_registry_generation {
+            format!("&{}=true", NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY)
+        } else {
+            String::new()
+        }
     )
 }
 
@@ -794,6 +888,7 @@ mod tests {
         let probe_err = transport
             .probe_ns_scanner(NsScannerCapabilityRequest {
                 endpoint: "http://node1:9000".to_string(),
+                supports_tier_registry_generation: false,
             })
             .await
             .expect_err("legacy transport should report namespace scanner as unsupported");
@@ -1387,6 +1482,7 @@ mod tests {
         let url = build_ns_scanner_capability_url(
             &NsScannerCapabilityRequest {
                 endpoint: "http://node1:9000".to_string(),
+                supports_tier_registry_generation: false,
             },
             challenge,
         );
@@ -1399,6 +1495,85 @@ mod tests {
         );
     }
 
+    #[test]
+    fn ns_scanner_capability_url_marks_generation_support_only_when_requested() {
+        let challenge = Uuid::new_v4();
+        let url = build_ns_scanner_capability_url(
+            &NsScannerCapabilityRequest {
+                endpoint: "http://node1:9000".to_string(),
+                supports_tier_registry_generation: true,
+            },
+            challenge,
+        );
+
+        assert!(url.contains(&format!("&{}=true", NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY)));
+    }
+
+    #[test]
+    fn ns_scanner_capability_legacy_fallback_requires_explicit_compatibility_status() {
+        for status in [400, 404, 405, 426] {
+            let error = Error::from(rustfs_rio::new_test_internode_http_io_error(
+                rustfs_rio::InternodeHttpErrorKind::HttpStatus(http::StatusCode::from_u16(status).expect("test status")),
+            ));
+            assert!(
+                ns_scanner_capability_error_allows_legacy(&error),
+                "status {status} should permit legacy retry"
+            );
+        }
+
+        let marked_server_error = Error::from(rustfs_rio::new_test_internode_http_io_error(
+            rustfs_rio::InternodeHttpErrorKind::HttpStatus(http::StatusCode::INTERNAL_SERVER_ERROR),
+        ));
+        let network_error = Error::from(rustfs_rio::new_test_internode_http_io_error(
+            rustfs_rio::InternodeHttpErrorKind::ConnectionRefused,
+        ));
+        let authentication_error = Error::other("remote namespace scanner capability authentication failed");
+        assert!(!ns_scanner_capability_error_allows_legacy(&marked_server_error));
+        assert!(!ns_scanner_capability_error_allows_legacy(&network_error));
+        assert!(!ns_scanner_capability_error_allows_legacy(&authentication_error));
+    }
+
+    #[test]
+    fn authenticated_ns_scanner_capability_ignores_unprotected_response_bit() {
+        let mut response = NsScannerCapabilityResponse {
+            version: NS_SCANNER_PROTOCOL_VERSION,
+            server_epoch: Uuid::new_v4(),
+            proof: Vec::new(),
+            supports_tier_registry_generation: None,
+        };
+
+        normalize_ns_scanner_capability_response(&mut response, true);
+        assert_eq!(response.supports_tier_registry_generation, Some(true));
+
+        response.supports_tier_registry_generation = Some(false);
+        normalize_ns_scanner_capability_response(&mut response, false);
+        assert_eq!(response.supports_tier_registry_generation, None);
+    }
+
+    #[test]
+    fn ns_scanner_capability_accepts_only_authenticated_legacy_scope_after_marker_mismatch() {
+        crate::runtime::sources::ensure_test_rpc_secret();
+        let challenge = Uuid::new_v4();
+        let response = NsScannerCapabilityResponse {
+            version: NS_SCANNER_PROTOCOL_VERSION,
+            server_epoch: Uuid::new_v4(),
+            proof: crate::cluster::rpc::sign_ns_scanner_capability(challenge, Uuid::new_v4())
+                .expect("placeholder proof should be generated"),
+            supports_tier_registry_generation: None,
+        };
+        // A proof bound to a different challenge cannot authorize the legacy
+        // fallback, even though the response has the expected shape.
+        assert!(!ns_scanner_capability_legacy_proof_is_valid(challenge, &response));
+
+        let server_epoch = response.server_epoch;
+        let valid_response = NsScannerCapabilityResponse {
+            proof: crate::cluster::rpc::sign_ns_scanner_capability(challenge, server_epoch)
+                .expect("legacy proof should be generated"),
+            ..response
+        };
+        assert!(ns_scanner_capability_legacy_proof_is_valid(challenge, &valid_response));
+    }
+
     #[test]
     fn transport_config_defaults_to_tcp_http() {
         let transport = build_internode_data_transport(None).unwrap();

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

@@ -35,8 +35,9 @@ pub use http_auth::{
     TONIC_RPC_PREFIX, build_auth_headers, build_put_file_auth_trailer, check_and_record_signed_rpc_nonce, gen_signature_headers,
     gen_tonic_replay_scope_headers, gen_tonic_signature_headers, normalize_tonic_rpc_audience, set_tonic_canonical_body_digest,
     set_tonic_mutation_body_digest, set_tonic_rolling_canonical_body_digest, set_tonic_rolling_mutation_body_digest,
-    sign_ns_scanner_capability, sign_put_file_capability, sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge,
-    tonic_boot_epoch_response_headers, tonic_rpc_auth_failure_reason, verify_ns_scanner_capability, verify_put_file_auth_trailer,
+    sign_ns_scanner_capability, sign_ns_scanner_capability_with_tier_registry_generation, sign_put_file_capability,
+    sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers, tonic_rpc_auth_failure_reason,
+    verify_ns_scanner_capability, verify_ns_scanner_capability_with_tier_registry_generation, verify_put_file_auth_trailer,
     verify_put_file_capability, verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
     verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
     verify_tonic_rpc_signature_with_bootstrap,

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

@@ -781,14 +781,16 @@ impl RemoteDisk {
         if self.health.is_faulty() {
             return Err(DiskError::FaultyDisk);
         }
-        let probe = self.data_transport.probe_ns_scanner(NsScannerCapabilityRequest {
+        let probe = self.data_transport.probe_ns_scanner_capability(NsScannerCapabilityRequest {
             endpoint: self.endpoint.grid_host(),
+            supports_tier_registry_generation: true,
         });
         let result = timeout(NS_SCANNER_CAPABILITY_PROBE_TIMEOUT, probe)
             .await
             .map_err(|_| DiskError::other("remote namespace scanner capability probe timed out"))?;
         match result {
-            Ok(server_epoch) => Ok(Some(server_epoch)),
+            Ok(response) if response.supports_tier_registry_generation == Some(true) => Ok(Some(response.server_epoch)),
+            Ok(_) => Ok(None),
             // RUSTFS_COMPAT_TODO(ns-scanner-rpc-v3): old peers and legacy transports lack the authenticated startup-epoch handshake. Remove after every supported peer implements namespace scanner protocol v3.
             Err(DiskError::MethodNotAllowed) => Ok(None),
             Err(err)
@@ -4040,10 +4042,21 @@ mod tests {
         NsScannerProbe(NsScannerCapabilityRequest),
     }
 
-    #[derive(Debug, Clone, Default)]
+    #[derive(Debug, Clone)]
     struct RecordingInternodeDataTransport {
         calls: Arc<StdMutex<Vec<RecordedTransportCall>>>,
         ns_scanner_probe_status: Arc<StdMutex<Option<u16>>>,
+        ns_scanner_generation_support: Arc<StdMutex<Option<bool>>>,
+    }
+
+    impl Default for RecordingInternodeDataTransport {
+        fn default() -> Self {
+            Self {
+                calls: Arc::default(),
+                ns_scanner_probe_status: Arc::default(),
+                ns_scanner_generation_support: Arc::new(StdMutex::new(Some(true))),
+            }
+        }
     }
 
     #[derive(Clone, Debug)]
@@ -4263,6 +4276,15 @@ mod tests {
             Self {
                 calls: Arc::default(),
                 ns_scanner_probe_status: Arc::new(StdMutex::new(Some(status))),
+                ns_scanner_generation_support: Arc::new(StdMutex::new(Some(true))),
+            }
+        }
+
+        fn with_ns_scanner_generation_support(support: Option<bool>) -> Self {
+            Self {
+                calls: Arc::default(),
+                ns_scanner_probe_status: Arc::default(),
+                ns_scanner_generation_support: Arc::new(StdMutex::new(support)),
             }
         }
 
@@ -4945,6 +4967,23 @@ mod tests {
             Ok(Uuid::from_u128(1))
         }
 
+        async fn probe_ns_scanner_capability(
+            &self,
+            request: NsScannerCapabilityRequest,
+        ) -> Result<crate::storage_api_contracts::internode::NsScannerCapabilityResponse> {
+            let server_epoch = self.probe_ns_scanner(request).await?;
+            let supports_tier_registry_generation = *self
+                .ns_scanner_generation_support
+                .lock()
+                .expect("namespace scanner generation support lock poisoned");
+            Ok(crate::storage_api_contracts::internode::NsScannerCapabilityResponse {
+                version: crate::storage_api_contracts::internode::NS_SCANNER_PROTOCOL_VERSION,
+                server_epoch,
+                proof: Vec::new(),
+                supports_tier_registry_generation,
+            })
+        }
+
         fn name(&self) -> &'static str {
             "recording"
         }
@@ -6757,6 +6796,22 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_remote_disk_namespace_scanner_capability_falls_back_without_generation_support() {
+        for support in [None, Some(false)] {
+            let transport = RecordingInternodeDataTransport::with_ns_scanner_generation_support(support);
+            let remote_disk = new_remote_disk_with_transport(Arc::new(transport)).await;
+
+            assert_eq!(
+                remote_disk
+                    .ns_scanner_server_epoch()
+                    .await
+                    .expect("missing generation support should be classified as unsupported"),
+                None
+            );
+        }
+    }
+
     #[tokio::test]
     async fn test_remote_disk_namespace_scanner_capability_rejects_legacy_transport() {
         let remote_disk = new_remote_disk_with_transport(Arc::new(RetryingOpenReadInternodeDataTransport::default())).await;

crates/ecstore/src/core/pools.rs

@@ -1455,36 +1455,6 @@ where
     Ok(())
 }
 
-async fn run_decommission_phases<F>(
-    rx: CancellationToken,
-    regular_buckets: Vec<DecomBucketInfo>,
-    meta_buckets: Vec<DecomBucketInfo>,
-    bucket_concurrency: usize,
-    mut start_bucket: F,
-) -> Result<()>
-where
-    F: FnMut(DecomBucketInfo, CancellationToken) -> BoxFuture<'static, Result<()>>,
-{
-    decommission_cancel_signal_result(rx.is_cancelled())?;
-
-    for bucket in meta_buckets {
-        decommission_cancel_signal_result(rx.is_cancelled())?;
-        start_bucket(bucket, rx.clone()).await?;
-    }
-
-    decommission_cancel_signal_result(rx.is_cancelled())?;
-
-    if bucket_concurrency <= 1 {
-        for bucket in regular_buckets {
-            decommission_cancel_signal_result(rx.is_cancelled())?;
-            start_bucket(bucket, rx.clone()).await?;
-        }
-        return Ok(());
-    }
-
-    run_decommission_buckets_bounded(rx, regular_buckets, bucket_concurrency, start_bucket).await
-}
-
 #[cfg(test)]
 async fn wait_decommission_worker_drain(workers: &Semaphore, limit: usize) -> Result<()> {
     let permits = u32::try_from(limit)
@@ -2056,7 +2026,7 @@ impl PoolMeta {
         self.load_no_lock(pool).await
     }
 
-    async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
+    pub(crate) async fn load_no_lock<S>(&mut self, pool: Arc<S>) -> Result<()>
     where
         S: EcstoreObjectIO,
     {
@@ -4932,6 +4902,25 @@ impl ECStore {
         Ok(())
     }
 
+    async fn decommission_buckets_concurrently(
+        self: &Arc<Self>,
+        rx: CancellationToken,
+        idx: usize,
+        pool: Arc<Sets>,
+        buckets: Vec<DecomBucketInfo>,
+        limit: usize,
+        entry_budget: Arc<Semaphore>,
+    ) -> Result<()> {
+        let store = Arc::clone(self);
+        run_decommission_buckets_bounded(rx, buckets, limit, move |bucket, rx| {
+            let store = Arc::clone(&store);
+            let pool = pool.clone();
+            let entry_budget = entry_budget.clone();
+            Box::pin(async move { store.decommission_pending_bucket(rx, idx, pool, bucket, entry_budget).await })
+        })
+        .await
+    }
+
     #[tracing::instrument(skip(self, rx))]
     async fn decommission_in_background(
         self: &Arc<Self>,
@@ -4946,15 +4935,31 @@ impl ECStore {
             pool_meta.pending_buckets(idx)
         };
         let bucket_concurrency = decommission_bucket_concurrency_limit();
+        if bucket_concurrency <= 1 {
+            for bucket in pending {
+                self.decommission_pending_bucket(rx.clone(), idx, pool.clone(), bucket, entry_budget.clone())
+                    .await?;
+            }
+            return Ok(());
+        }
+
         let (regular_buckets, meta_buckets) = split_decommission_buckets(pending);
-        let store = Arc::clone(self);
-        run_decommission_phases(rx, regular_buckets, meta_buckets, bucket_concurrency, move |bucket, rx| {
-            let store = Arc::clone(&store);
-            let pool = pool.clone();
-            let entry_budget = entry_budget.clone();
-            Box::pin(async move { store.decommission_pending_bucket(rx, idx, pool, bucket, entry_budget).await })
-        })
-        .await
+        self.decommission_buckets_concurrently(
+            rx.clone(),
+            idx,
+            pool.clone(),
+            regular_buckets,
+            bucket_concurrency,
+            entry_budget.clone(),
+        )
+        .await?;
+
+        for bucket in meta_buckets {
+            self.decommission_pending_bucket(rx.clone(), idx, pool.clone(), bucket, entry_budget.clone())
+                .await?;
+        }
+
+        Ok(())
     }
 
     #[tracing::instrument(skip(self))]
@@ -6372,8 +6377,8 @@ mod pools_tests {
         resolve_decommission_terminal_mark_after_error_result, resolve_decommission_terminal_mark_result,
         resolve_decommission_update_after_result, resolve_start_decommission_pool_meta_reload_result,
         rollback_start_decommission_pool_meta, run_decommission_buckets_bounded, run_decommission_listing_with_retry,
-        run_decommission_listing_with_retry_and_drain, run_decommission_phases, run_decommission_side_effect,
-        should_cleanup_decommission_source_entry, should_continue_decommission_queue, should_count_decommission_version_complete,
+        run_decommission_listing_with_retry_and_drain, run_decommission_side_effect, should_cleanup_decommission_source_entry,
+        should_continue_decommission_queue, should_count_decommission_version_complete,
         should_preserve_decommission_canceled_state, should_reject_decommission_cancel_as_terminal,
         should_retry_decommission_cancel_reload, should_retry_decommission_listing, should_skip_canceled_decommission_routine,
         spawn_decommission_index_cancelers, split_decommission_buckets, take_and_cancel_decommission_canceler,
@@ -6392,7 +6397,7 @@ mod pools_tests {
     use rustfs_filemeta::{MetaCacheEntries, MetadataResolutionParams};
     use rustfs_rio::Index;
     use std::sync::{
-        Arc, Mutex as StdMutex,
+        Arc,
         atomic::{AtomicBool, AtomicUsize, Ordering},
     };
     use std::time::Duration as StdDuration;
@@ -6685,66 +6690,6 @@ mod pools_tests {
         );
     }
 
-    #[tokio::test]
-    async fn test_decommission_metadata_phase_precedes_regular_failure() {
-        let events = Arc::new(StdMutex::new(Vec::new()));
-        let err = run_decommission_phases(
-            CancellationToken::new(),
-            vec![
-                DecomBucketInfo {
-                    name: "regular-fails".to_string(),
-                    ..Default::default()
-                },
-                DecomBucketInfo {
-                    name: "regular-not-started".to_string(),
-                    ..Default::default()
-                },
-            ],
-            vec![
-                DecomBucketInfo {
-                    name: crate::disk::RUSTFS_META_BUCKET.to_string(),
-                    prefix: crate::config::com::CONFIG_PREFIX.to_string(),
-                },
-                DecomBucketInfo {
-                    name: crate::disk::RUSTFS_META_BUCKET.to_string(),
-                    prefix: crate::disk::BUCKET_META_PREFIX.to_string(),
-                },
-            ],
-            1,
-            {
-                let events = Arc::clone(&events);
-                move |bucket, _rx| {
-                    let events = Arc::clone(&events);
-                    Box::pin(async move {
-                        let event = if bucket.name == crate::disk::RUSTFS_META_BUCKET {
-                            format!("meta:{}", bucket.prefix)
-                        } else {
-                            format!("regular:{}", bucket.name)
-                        };
-                        events.lock().expect("phase event lock should not be poisoned").push(event);
-                        if bucket.name == "regular-fails" {
-                            Err(Error::SlowDown)
-                        } else {
-                            Ok(())
-                        }
-                    })
-                }
-            },
-        )
-        .await
-        .expect_err("regular failure should remain fatal after metadata completes");
-
-        assert!(matches!(err, Error::SlowDown));
-        assert_eq!(
-            *events.lock().expect("phase event lock should not be poisoned"),
-            vec![
-                format!("meta:{}", crate::config::com::CONFIG_PREFIX),
-                format!("meta:{}", crate::disk::BUCKET_META_PREFIX),
-                "regular:regular-fails".to_string(),
-            ]
-        );
-    }
-
     #[tokio::test]
     async fn test_run_decommission_buckets_bounded_respects_limit() {
         let rx = CancellationToken::new();

crates/ecstore/src/core/sets.rs

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

crates/ecstore/src/io_support/bitrot.rs

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

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

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

crates/ecstore/src/storage_api_contracts/mod.rs

@@ -27,12 +27,12 @@ pub(crate) mod internode {
         NS_SCANNER_BODY_SHA256_QUERY, NS_SCANNER_CAPABILITY_CHALLENGE_QUERY, NS_SCANNER_CYCLE_QUERY,
         NS_SCANNER_LEADER_EPOCH_QUERY, NS_SCANNER_PROTOCOL_VERSION, NS_SCANNER_PROTOCOL_VERSION_QUERY,
         NS_SCANNER_REQUEST_ID_QUERY, NS_SCANNER_SERVER_EPOCH_QUERY, NS_SCANNER_SESSION_ID_QUERY,
-        NS_SCANNER_SESSION_SEQUENCE_QUERY, NsScannerCapabilityResponse, PUT_FILE_AUTH_QUERY, PUT_FILE_AUTH_TRAILER_DIGEST_LEN,
-        PUT_FILE_AUTH_TRAILER_LEN, PUT_FILE_AUTH_TRAILER_MAC_LEN, PUT_FILE_AUTH_TRAILER_MAGIC, PUT_FILE_AUTH_V1,
-        PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY, PUT_FILE_CAPABILITY_VERSION, PUT_FILE_NONCE_QUERY,
-        PUT_FILE_SERVER_EPOCH_QUERY, PutFileCapabilityResponse, SCANNER_ACTIVITY_LEGACY_PROTOCOL_VERSION,
-        SCANNER_ACTIVITY_PREVIOUS_PROTOCOL_VERSION, SCANNER_ACTIVITY_PROTOCOL_VERSION, WALK_DIR_BODY_SHA256_QUERY,
-        WALK_DIR_STREAM_COMPLETION_QUERY, WALK_DIR_STREAM_COMPLETION_V1,
+        NS_SCANNER_SESSION_SEQUENCE_QUERY, NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY, NsScannerCapabilityResponse,
+        PUT_FILE_AUTH_QUERY, PUT_FILE_AUTH_TRAILER_DIGEST_LEN, PUT_FILE_AUTH_TRAILER_LEN, PUT_FILE_AUTH_TRAILER_MAC_LEN,
+        PUT_FILE_AUTH_TRAILER_MAGIC, PUT_FILE_AUTH_V1, PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY,
+        PUT_FILE_CAPABILITY_VERSION, PUT_FILE_NONCE_QUERY, PUT_FILE_SERVER_EPOCH_QUERY, PutFileCapabilityResponse,
+        SCANNER_ACTIVITY_LEGACY_PROTOCOL_VERSION, SCANNER_ACTIVITY_PREVIOUS_PROTOCOL_VERSION, SCANNER_ACTIVITY_PROTOCOL_VERSION,
+        WALK_DIR_BODY_SHA256_QUERY, WALK_DIR_STREAM_COMPLETION_QUERY, WALK_DIR_STREAM_COMPLETION_V1,
     };
 }
 

crates/ecstore/src/store/heal.rs

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

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

@@ -231,6 +231,10 @@ impl HealTask {
                             "Heal erasure set format repair skipped because no format heal was required"
                         );
                     } else {
+                        let error = e;
+                        if error.is_recoverable_heal() {
+                            return Err(error);
+                        }
                         error!(
                             target: "rustfs::heal::task",
                             event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -239,7 +243,7 @@ impl HealTask {
                             task_id = %self.id,
                             set_disk_id,
                             result = "format_failed",
-                            error = %e,
+                            error = %error,
                             "Heal erasure set failed"
                         );
                         {
@@ -247,7 +251,7 @@ impl HealTask {
                             progress.update_progress(4, 4, 0, 0);
                         }
                         return Err(Error::TaskExecutionFailed {
-                            message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
+                            message: format!("Failed to heal disk format for {set_disk_id}: {error}"),
                         });
                     }
                 } else {
@@ -284,6 +288,9 @@ impl HealTask {
             Err(Error::TaskCancelled) => return Err(Error::TaskCancelled),
             Err(Error::TaskTimeout) => return Err(Error::TaskTimeout),
             Err(e) => {
+                if e.is_recoverable_heal() {
+                    return Err(e);
+                }
                 error!(
                     target: "rustfs::heal::task",
                     event = EVENT_HEAL_ERASURE_SET_RESULT,

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

@@ -547,6 +547,7 @@ struct MockStorage {
     heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
     heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
     format_no_heal_required: Mutex<bool>,
+    format_error: Mutex<Option<Error>>,
     global_format_calls: Mutex<u32>,
     replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
     replacement_targets_ready: Mutex<bool>,
@@ -867,6 +868,9 @@ impl HealStorageAPI for MockStorage {
 
     async fn heal_format(&self, _dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
         *self.global_format_calls.lock().unwrap() += 1;
+        if let Some(error) = self.format_error.lock().unwrap().take() {
+            return Err(error);
+        }
         let no_heal_required = *self.format_no_heal_required.lock().unwrap();
         if no_heal_required {
             Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
@@ -2052,6 +2056,30 @@ async fn test_erasure_set_heal_continues_after_format_no_heal_required() {
     );
 }
 
+#[tokio::test]
+async fn erasure_set_format_slowdown_is_propagated() {
+    let storage = Arc::new(MockStorage {
+        format_error: Mutex::new(Some(Error::Storage(EcstoreError::SlowDown))),
+        ..Default::default()
+    });
+    let request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions::default(),
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("format SlowDown must remain recoverable for the task manager");
+
+    assert!(matches!(error, Error::Storage(EcstoreError::SlowDown)));
+}
+
 #[tokio::test]
 async fn erasure_set_bucket_prepass_failure_stops_before_object_heal() {
     let temp = TempDir::new().expect("temporary directory should be created");

crates/kms/tests/vault_ha_failover_live.rs

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

crates/scanner/src/data_usage_define.rs

@@ -29,7 +29,8 @@ use rustfs_config::ENV_SCANNER_CACHE_SAVE_TIMEOUT_SECS;
 pub use rustfs_data_usage::{
     AllTierStats, BucketTargetUsageInfo, BucketUsageInfo, DATA_USAGE_OBJECT_NAME, DATA_USAGE_OBSERVED_OBJECT_NAME,
     DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, LEGACY_DATA_USAGE_OBJECT_NAME, PrefixUsageEntry,
-    PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeSummary, TierStats, hash_path, prefix_usage_in_cache,
+    PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeSummary, TierAccountingProof, TierStats, UNKNOWN_TIER,
+    UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP, UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP, UnknownTierStats, hash_path, prefix_usage_in_cache,
 };
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
 use tokio::time::{Duration, Instant, sleep, timeout};
@@ -205,25 +206,79 @@ impl ScannerSizeSummaryExt for SizeSummary {
             self.versions = self.versions.saturating_add(1);
         }
 
-        let size = usize::try_from(size.max(0)).unwrap_or(usize::MAX);
+        let logical_size = size.max(0);
+        let size = usize::try_from(logical_size).unwrap_or(usize::MAX);
         self.total_size = self.total_size.saturating_add(size);
+        let logical_bytes = u64::try_from(logical_size).unwrap_or(u64::MAX);
+        let physical_bytes = u64::try_from(oi.size.max(0)).unwrap_or(0);
+        let mut proof = TierAccountingProof {
+            logical_total: logical_bytes,
+            logical_known: 0,
+            physical_total: physical_bytes,
+            physical_known: 0,
+            overflowed: false,
+        };
 
         if oi.transitioned_object.free_version {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
+            self.tier_accounting_proof.saturating_add(proof);
             return;
         }
 
-        let mut tier = oi.storage_class.clone().unwrap_or_else(|| storageclass::STANDARD.to_string());
-        if oi.transitioned_object.status == TRANSITION_COMPLETE {
-            tier = oi.transitioned_object.tier.clone();
+        let tier = if oi.transitioned_object.status == TRANSITION_COMPLETE {
+            oi.transitioned_object.tier.as_str()
+        } else {
+            oi.storage_class.as_deref().unwrap_or(storageclass::STANDARD)
+        };
+
+        let builtin_tier = tier == storageclass::STANDARD || tier == storageclass::RRS;
+        let tier_registry_is_empty =
+            self.tier_stats.is_empty() || (self.tier_stats.len() == 1 && self.tier_stats.contains_key(UNKNOWN_TIER));
+        let known_tier = tier != UNKNOWN_TIER && (builtin_tier || self.tier_stats.contains_key(tier));
+
+        // With no configured tier, retain the historical empty-map shape for
+        // ordinary STANDARD/RRS objects. A non-built-in key is still an
+        // observable unknown and must create only the fixed bucket.
+        if tier_registry_is_empty && known_tier {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
+            self.tier_accounting_proof.saturating_add(proof);
+            return;
         }
 
-        if let Some(tier_stats) = self.tier_stats.get_mut(&tier) {
-            *tier_stats = tier_stats.add(&TierStats {
-                total_size: u64::try_from(oi.size).unwrap_or(0),
-                num_versions: 1,
-                num_objects: u64::from(oi.is_latest),
-            });
+        // Configured tiers and the fixed bucket are normally seeded, so the
+        // hot path can mutate them without allocating a key for every object.
+        // The fallback inserts only when a legacy/no-config summary sees its
+        // first unknown key.
+        let tier_stats = if known_tier {
+            if let Some(stats) = self.tier_stats.get_mut(tier) {
+                stats
+            } else {
+                self.tier_stats.entry(tier.to_owned()).or_default()
+            }
+        } else if let Some(stats) = self.tier_stats.get_mut(UNKNOWN_TIER) {
+            stats
+        } else {
+            self.tier_stats.entry(UNKNOWN_TIER.to_string()).or_default()
+        };
+        *tier_stats = tier_stats.add(&TierStats {
+            total_size: physical_bytes,
+            num_versions: 1,
+            num_objects: u64::from(oi.is_latest),
+        });
+        if known_tier {
+            proof.logical_known = logical_bytes;
+            proof.physical_known = physical_bytes;
         }
+        if !known_tier {
+            self.unknown_tier_stats
+                .record_dimensions(tier, logical_bytes, physical_bytes, 1, u64::from(oi.is_latest));
+            if self.unknown_tier_stats.counter_overflowed {
+                proof.overflowed = true;
+            }
+        }
+        self.tier_accounting_proof.saturating_add(proof);
     }
 }
 
@@ -271,6 +326,11 @@ pub struct DataUsageEntryInfo {
     pub name: String,
     pub parent: String,
     pub entry: DataUsageEntry,
+    /// Registry generation used to classify this root entry. Older remote
+    /// workers omit it; callers must reject that result when a frozen cycle
+    /// requires generation fencing.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tier_registry_generation: Option<u64>,
 }
 
 #[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq, Hash)]
@@ -344,6 +404,10 @@ pub struct DataUsageCacheInfo {
     pub scan_plan_digest: Option<DataUsageScanPlanDigest>,
     #[serde(default)]
     pub cache_key_format: u16,
+    /// Registry generation used for the completed/partial scan. This is
+    /// process-local audit data; older cache writers omit it.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tier_registry_generation: Option<u64>,
 }
 
 impl Serialize for DataUsageCacheInfo {
@@ -353,7 +417,8 @@ impl Serialize for DataUsageCacheInfo {
     {
         // Keep this metadata map-encoded so older readers can ignore fields
         // appended by newer scanner versions during rolling upgrades.
-        let mut state = serializer.serialize_map(Some(16))?;
+        let field_count = 16 + usize::from(self.tier_registry_generation.is_some());
+        let mut state = serializer.serialize_map(Some(field_count))?;
         state.serialize_entry("name", &self.name)?;
         state.serialize_entry("next_cycle", &self.next_cycle)?;
         state.serialize_entry("leader_epoch", &self.leader_epoch)?;
@@ -370,6 +435,9 @@ impl Serialize for DataUsageCacheInfo {
         state.serialize_entry("snapshot_complete", &self.snapshot_complete)?;
         state.serialize_entry("scan_plan_digest", &self.scan_plan_digest)?;
         state.serialize_entry("cache_key_format", &self.cache_key_format)?;
+        if let Some(generation) = self.tier_registry_generation {
+            state.serialize_entry("tier_registry_generation", &generation)?;
+        }
         state.end()
     }
 }
@@ -390,6 +458,58 @@ pub(crate) enum DataUsageCachePrepareOutcome {
 }
 
 impl DataUsageCache {
+    /// Reconcile tier keys loaded from an older cache against the registry
+    /// frozen for this scan. New metadata is already routed through
+    /// `UNKNOWN_TIER`; this pass handles retired keys that predate that rule.
+    /// Legacy `TierStats` carries physical bytes only, so this migration does
+    /// not manufacture a logical unknown-byte value from that physical total.
+    pub(crate) fn fold_retired_tiers(&mut self, tier_names: &[String]) {
+        let known_tiers = tier_names.iter().map(String::as_str).collect::<HashSet<_>>();
+        for entry in self.cache.values_mut() {
+            let Some(tiers) = entry.all_tier_stats.as_mut() else { continue };
+            let existing_unknown = tiers.tiers.get(UNKNOWN_TIER).cloned().unwrap_or_default();
+            let companion_present = entry.unknown_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty());
+            let migrate_existing_unknown = !companion_present;
+            let mut retired = TierStats::default();
+            let mut retired_key_found = false;
+            if migrate_existing_unknown {
+                retired = retired.add(&existing_unknown);
+            }
+            for (tier, stats) in &tiers.tiers {
+                if tier != UNKNOWN_TIER
+                    && tier != storageclass::STANDARD
+                    && tier != storageclass::RRS
+                    && !known_tiers.contains(tier.as_str())
+                {
+                    retired_key_found = true;
+                    retired = retired.add(stats);
+                }
+            }
+            tiers.fold_unknown_tiers(tier_names.iter().map(String::as_str));
+            if !retired.is_empty() && !companion_present {
+                entry.add_unknown_tier_stats(&UnknownTierStats {
+                    // The legacy map stores physical bytes only. Logical
+                    // bytes remain zero until a fresh object scan observes
+                    // them under the current metadata format.
+                    unknown_physical_bytes: retired.total_size,
+                    unknown_objects: retired.num_objects,
+                    unknown_versions: retired.num_versions,
+                    ..Default::default()
+                });
+                // The legacy tier map has no logical-byte dimension, so a
+                // proof that classified this retired key as known cannot be
+                // repaired safely. Mark it unvalidated and require a fresh
+                // scan rather than guessing a logical subtraction.
+                entry.tier_accounting_proof = None;
+            } else if retired_key_found {
+                // A nonempty companion has no provenance tying it to the
+                // retired map keys. Reject the mixed cache until a fresh scan
+                // reconciles the dimensions instead of double-counting them.
+                entry.tier_accounting_proof = None;
+            }
+        }
+    }
+
     /// Prefix-level usage query over this (writer-side) cache; see
     /// [`prefix_usage_in_cache`] for the semantics
     /// (rustfs/backlog#1872).
@@ -875,6 +995,7 @@ impl DataUsageCache {
             delete_markers_total_count: flat.delete_markers as u64,
             objects_total_size: flat.size as u64,
             tier_stats: flat.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+            unknown_tier_stats: flat.unknown_tier_stats.filter(|stats| !stats.is_empty()),
             buckets_count: u64::try_from(buckets.len()).unwrap_or(u64::MAX),
             buckets_usage,
             ..Default::default()

crates/scanner/src/data_usage_define/tests.rs

@@ -573,7 +573,6 @@ fn size_summary_add_saturates_all_usage_counters() {
             failed_count: usize::MAX,
         },
     );
-
     let mut increment = SizeSummary {
         total_size: 1,
         versions: 1,
@@ -588,6 +587,24 @@ fn size_summary_add_saturates_all_usage_counters() {
         failed_count: 1,
         ..Default::default()
     };
+    summary.tier_stats.insert(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: u64::MAX,
+            num_versions: u64::MAX,
+            num_objects: u64::MAX,
+        },
+    );
+    increment.tier_stats.insert(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 1,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    );
+    increment.unknown_tier_stats.unknown_bytes = 1;
+    increment.unknown_tier_stats.unknown_physical_bytes = 1;
     increment.repl_target_stats.insert(
         target.clone(),
         ReplTargetSizeSummary {
@@ -624,6 +641,8 @@ fn size_summary_add_saturates_all_usage_counters() {
     assert_eq!(target_summary.failed_size, i64::MAX);
     assert_eq!(target_summary.pending_count, usize::MAX);
     assert_eq!(target_summary.failed_count, usize::MAX);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, u64::MAX);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 1);
 }
 
 #[test]
@@ -673,6 +692,249 @@ fn size_summary_actions_accounting_accumulates_tier_stats() {
     );
 }
 
+#[test]
+fn unknown_tier_is_bounded_and_accounted() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    let object = ObjectInfo {
+        storage_class: Some("retired-tier".to_string()),
+        size: 11,
+        is_latest: true,
+        ..Default::default()
+    };
+
+    summary.actions_accounting(&object, 11, 11);
+
+    assert_eq!(summary.tier_stats.len(), 2);
+    assert_eq!(summary.tier_stats.get(UNKNOWN_TIER).map(|stats| stats.total_size), Some(11));
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 11);
+    assert_eq!(summary.unknown_tier_stats.unknown_physical_bytes, 11);
+    assert_eq!(summary.unknown_tier_stats.unknown_objects, 1);
+    assert_eq!(summary.tier_accounting_proof.logical_total, 11);
+    assert_eq!(summary.tier_accounting_proof.logical_known, 0);
+    assert_eq!(summary.tier_accounting_proof.physical_total, 11);
+    assert_eq!(summary.tier_accounting_proof.physical_known, 0);
+    assert!(summary.unknown_tier_stats.diagnostics.len() <= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP);
+    assert!(summary.unknown_tier_stats.diagnostics.iter().map(String::len).sum::<usize>() <= UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP);
+    assert!(
+        summary
+            .unknown_tier_stats
+            .diagnostics
+            .iter()
+            .all(|entry| !entry.contains("retired"))
+    );
+}
+
+#[test]
+fn unknown_tier_is_accounted_when_no_remote_tier_is_configured() {
+    let mut summary = SizeSummary::new();
+    let object = ObjectInfo {
+        storage_class: Some("retired-tier".to_string()),
+        size: 3,
+        is_latest: true,
+        ..Default::default()
+    };
+
+    summary.actions_accounting(&object, 9, 9);
+
+    assert_eq!(summary.tier_stats.len(), 1);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 3);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 9);
+    assert_eq!(summary.unknown_tier_stats.unknown_physical_bytes, 3);
+
+    let standard = ObjectInfo {
+        storage_class: Some(storageclass::STANDARD.to_string()),
+        size: 4,
+        is_latest: true,
+        ..Default::default()
+    };
+    summary.actions_accounting(&standard, 4, 4);
+    assert_eq!(summary.tier_accounting_proof.logical_total, 13);
+    assert_eq!(summary.tier_accounting_proof.logical_known, 4);
+    assert_eq!(summary.tier_accounting_proof.physical_total, 3 + 4);
+    assert_eq!(summary.tier_accounting_proof.physical_known, 4);
+    assert_eq!(summary.tier_stats.len(), 1, "built-ins preserve the no-tier map shape");
+}
+
+#[test]
+fn million_unique_tier_keys_do_not_grow_stats_map() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    for index in 0..1_000_000_u64 {
+        let object = ObjectInfo {
+            storage_class: Some(format!("untrusted-tier-{index}")),
+            size: 1,
+            ..Default::default()
+        };
+        summary.actions_accounting(&object, 1, 1);
+    }
+
+    assert_eq!(summary.tier_stats.len(), 2);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 1_000_000);
+    assert_eq!(summary.unknown_tier_stats.unknown_bytes, 1_000_000);
+    assert!(summary.unknown_tier_stats.diagnostics.len() <= UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP);
+    assert!(summary.unknown_tier_stats.diagnostics.iter().map(String::len).sum::<usize>() <= UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP);
+}
+
+#[test]
+fn unknown_tier_never_triggers_transition() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("WARM".to_string(), TierStats::default());
+    let mut object = ObjectInfo {
+        storage_class: Some("removed-tier".to_string()),
+        size: 7,
+        ..Default::default()
+    };
+    object.transitioned_object.status = TRANSITION_COMPLETE.to_string();
+    object.transitioned_object.tier = "removed-tier".to_string();
+
+    summary.actions_accounting(&object, 7, 7);
+
+    assert_eq!(summary.tier_stats.get("removed-tier"), None);
+    assert_eq!(summary.tier_stats[UNKNOWN_TIER].total_size, 7);
+}
+
+#[test]
+fn removed_tier_survives_restart_as_unknown() {
+    let mut summary = SizeSummary::new();
+    summary.tier_stats.insert("COLD".to_string(), TierStats::default());
+    summary.tier_stats.insert(
+        "RETIRED".to_string(),
+        TierStats {
+            total_size: 5,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    );
+    let object = ObjectInfo {
+        storage_class: Some("COLD".to_string()),
+        size: 5,
+        ..Default::default()
+    };
+    summary.actions_accounting(&object, 5, 5);
+    let mut entry = DataUsageEntry::default();
+    entry.add_tier_sizes(&summary.tier_stats);
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 2,
+        unknown_physical_bytes: 2,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    let encoded = rmp_serde::to_vec(&entry).expect("entry should encode");
+    let restored: DataUsageEntry = rmp_serde::from_slice(&encoded).expect("entry should decode");
+    assert_eq!(restored.unknown_tier_stats.as_ref().map(|stats| stats.unknown_bytes), Some(2));
+    assert_eq!(
+        restored.all_tier_stats.as_ref().expect("tier stats persisted").tiers["RETIRED"].total_size,
+        5
+    );
+
+    let mut cache = DataUsageCache::default();
+    cache.replace("bucket", "", restored);
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let folded = cache.cache.get(&hash_path("bucket").key()).expect("folded cache entry");
+    assert_eq!(
+        folded.all_tier_stats.as_ref().expect("tier stats persisted").tiers[UNKNOWN_TIER].total_size,
+        5
+    );
+    assert_eq!(folded.unknown_tier_stats.as_ref().map(|stats| stats.unknown_bytes), Some(2));
+}
+
+#[test]
+fn retired_tier_fold_is_idempotent_and_rejects_mixed_companion_provenance() {
+    let mut cache = DataUsageCache::default();
+    let mut entry = DataUsageEntry {
+        all_tier_stats: Some(AllTierStats {
+            tiers: HashMap::from([(
+                "RETIRED".to_string(),
+                TierStats {
+                    total_size: 5,
+                    num_versions: 1,
+                    num_objects: 1,
+                },
+            )]),
+        }),
+        ..Default::default()
+    };
+    entry.unknown_tier_stats = Some(UnknownTierStats {
+        unknown_physical_bytes: 5,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 5,
+        physical_known: 5,
+        ..Default::default()
+    });
+    cache.replace("bucket", "", entry);
+
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let first = cache.cache.get(&hash_path("bucket").key()).expect("entry").clone();
+    assert_eq!(first.all_tier_stats.as_ref().expect("tiers").tiers[UNKNOWN_TIER].total_size, 5);
+    assert_eq!(first.unknown_tier_stats.as_ref().expect("companion").unknown_physical_bytes, 5);
+    assert!(first.tier_accounting_proof.is_none(), "mixed provenance must not publish");
+
+    cache.fold_retired_tiers(&["COLD".to_string()]);
+    let second = cache.cache.get(&hash_path("bucket").key()).expect("entry");
+    assert_eq!(second.all_tier_stats.as_ref().expect("tiers").tiers[UNKNOWN_TIER].total_size, 5);
+    assert_eq!(second.unknown_tier_stats.as_ref().expect("companion").unknown_physical_bytes, 5);
+}
+
+#[test]
+fn tier_registry_refresh_does_not_mix_cycle_generations() {
+    let first = crate::TierRegistrySnapshot {
+        generation: 1,
+        names: Arc::from(["WARM".to_string()]),
+        refresh_failed: false,
+    };
+    let second = crate::TierRegistrySnapshot {
+        generation: 2,
+        names: Arc::from(["COLD".to_string()]),
+        refresh_failed: false,
+    };
+    assert_ne!(first.generation, second.generation);
+    assert_eq!(first.names.as_ref(), ["WARM".to_string()]);
+    assert_eq!(second.names.as_ref(), ["COLD".to_string()]);
+    assert!(first.refreshed(Err(())).refresh_failed);
+    assert!(!second.refreshed(Ok(Arc::from(["HOT".to_string()]))).refresh_failed);
+    assert_eq!(first.refreshed(Err(())).generation, first.generation);
+    assert_eq!(first.refreshed(Err(())).names, first.names);
+}
+
+#[test]
+fn unknown_tier_counter_uses_checked_arithmetic() {
+    let max = TierStats {
+        total_size: u64::MAX,
+        num_versions: u64::MAX,
+        num_objects: u64::MAX,
+    };
+    assert!(max.checked_add(&TierStats::default()).is_some());
+    assert!(
+        max.checked_add(&TierStats {
+            total_size: 1,
+            ..Default::default()
+        })
+        .is_none()
+    );
+
+    let mut unknown = UnknownTierStats {
+        unknown_bytes: u64::MAX,
+        ..Default::default()
+    };
+    unknown.record("overflow", 1, 1, 1);
+    assert_eq!(unknown.unknown_bytes, u64::MAX);
+    assert_eq!(unknown.unknown_objects, 1);
+    assert!(unknown.counter_overflowed);
+    assert!(unknown.checked_add(&UnknownTierStats::default()).is_none());
+    assert!(
+        unknown
+            .checked_add(&UnknownTierStats {
+                unknown_bytes: 1,
+                ..Default::default()
+            })
+            .is_none()
+    );
+}
+
 #[test]
 fn test_data_usage_entry_merge_sums_failed_objects() {
     let mut left = DataUsageEntry {

crates/scanner/src/lib.rs

@@ -24,8 +24,11 @@
 use bytes::Bytes;
 use http::HeaderMap;
 use rustfs_config::server_config::{Config as ServerConfig, get_global_server_config as config_get_global_server_config};
+use sha2::{Digest as _, Sha256};
+use std::collections::HashMap;
 use std::path::PathBuf;
 use std::sync::Arc;
+use std::sync::LazyLock;
 use std::sync::RwLock;
 use std::time::{Duration, Instant};
 use storage_api::owner::{
@@ -94,6 +97,45 @@ static SCANNER_RUNTIME_INSTANCES: AtomicU64 = AtomicU64::new(0);
 static SCANNER_FOREGROUND_READ_ACTIVITY: AtomicU64 = AtomicU64::new(0);
 static SCANNER_FOREGROUND_STREAM_READS: AtomicU64 = AtomicU64::new(0);
 
+/// Immutable tier registry captured at the beginning of a folder scan.
+/// Generation makes it possible to prove that a result was classified against
+/// one registry even when the process-wide TTL cache refreshes concurrently.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) struct TierRegistrySnapshot {
+    pub(crate) generation: u64,
+    pub(crate) names: Arc<[String]>,
+    /// True when the last refresh attempt failed and `names` is therefore a
+    /// retained last-good snapshot rather than a newly read registry.
+    pub(crate) refresh_failed: bool,
+}
+
+impl TierRegistrySnapshot {
+    /// Apply a refresh only when the registry read succeeds. A failed refresh
+    /// retains the prior generation, preventing a transient config failure
+    /// from classifying the remainder of a scan against an empty registry.
+    pub(crate) fn refreshed(&self, names: Result<Arc<[String]>, ()>) -> Self {
+        match names {
+            Ok(names) => Self {
+                generation: self.generation.saturating_add(1),
+                names,
+                refresh_failed: false,
+            },
+            Err(()) => Self {
+                refresh_failed: true,
+                ..self.clone()
+            },
+        }
+    }
+
+    pub(crate) fn initial(names: Arc<[String]>) -> Self {
+        Self {
+            generation: 1,
+            names,
+            refresh_failed: false,
+        }
+    }
+}
+
 pub fn current_scanner_activity() -> u64 {
     SCANNER_ACTIVE_WORK_UNITS.load(Ordering::Relaxed)
 }
@@ -371,6 +413,7 @@ pub(crate) fn resolve_scanner_server_config() -> Option<ServerConfig> {
 /// How long the scanner caches the runtime tier-name list before re-reading
 /// the tier configuration manager.
 const TIER_NAME_CACHE_TTL: Duration = Duration::from_secs(30);
+const MAX_TIER_REGISTRY_NAME_BYTES: usize = 256;
 
 /// Process-wide TTL cache of runtime tier names.
 ///
@@ -383,24 +426,192 @@ const TIER_NAME_CACHE_TTL: Duration = Duration::from_secs(30);
 /// `TIER_NAME_CACHE_TTL` later; a removed tier can leave an all-zero
 /// `TierStats` seed behind for one cache generation, which merges harmlessly
 /// by key in per-object accounting and disappears on the next refresh.
-static TIER_NAME_CACHE: RwLock<Option<(Instant, Arc<[String]>)>> = RwLock::new(None);
+static TIER_NAME_CACHE: RwLock<Option<(Instant, TierRegistrySnapshot)>> = RwLock::new(None);
+static TIER_REGISTRY_GENERATION: AtomicU64 = AtomicU64::new(0);
+static TIER_CYCLE_SNAPSHOTS: LazyLock<RwLock<HashMap<(u64, u64), TierRegistrySnapshot>>> =
+    LazyLock::new(|| RwLock::new(HashMap::new()));
+static TIER_ACTIVE_CYCLES: LazyLock<RwLock<HashMap<(u64, u64), usize>>> = LazyLock::new(|| RwLock::new(HashMap::new()));
+static TIER_NAME_REFRESH_LOCK: LazyLock<tokio::sync::Mutex<()>> = LazyLock::new(|| tokio::sync::Mutex::new(()));
+
+/// Return one immutable registry snapshot for a scanner unit of work.
+pub(crate) async fn runtime_tier_registry() -> TierRegistrySnapshot {
+    {
+        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
+        if let Some((refreshed_at, snapshot)) = cached
+            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
+        {
+            return snapshot;
+        }
+    }
+
+    // Serialize refreshes so a slower read of the old config cannot overwrite
+    // a newer snapshot published by a concurrent caller.
+    let _refresh_guard = TIER_NAME_REFRESH_LOCK.lock().await;
+    {
+        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
+        if let Some((refreshed_at, snapshot)) = cached
+            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
+        {
+            return snapshot;
+        }
+    }
+
+    let previous = TIER_NAME_CACHE
+        .read()
+        .unwrap_or_else(|err| err.into_inner())
+        .as_ref()
+        .map(|(_, snapshot)| snapshot.clone());
+    let names = ecstore_get_global_tier_config_mgr()
+        .read()
+        .await
+        .list_tiers()
+        .into_iter()
+        .map(|tier| tier.name)
+        .collect::<Vec<_>>();
+    let snapshot = match validate_tier_registry_names(names) {
+        Ok(names) => {
+            let generation = next_tier_registry_generation();
+            match previous {
+                Some(previous) => TierRegistrySnapshot {
+                    generation,
+                    ..previous.refreshed(Ok(names))
+                },
+                None => TierRegistrySnapshot {
+                    generation,
+                    ..TierRegistrySnapshot::initial(names)
+                },
+            }
+        }
+        Err(()) => match previous {
+            Some(previous) => previous.refreshed(Err(())),
+            None => TierRegistrySnapshot {
+                generation: next_tier_registry_generation(),
+                names: Arc::new([]),
+                refresh_failed: true,
+            },
+        },
+    };
+    *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((Instant::now(), snapshot.clone()));
+    snapshot
+}
+
+fn next_tier_registry_generation() -> u64 {
+    TIER_REGISTRY_GENERATION
+        .fetch_update(Ordering::AcqRel, Ordering::Relaxed, |current| Some(current.saturating_add(1)))
+        .unwrap_or(u64::MAX)
+}
+
+fn validate_tier_registry_names(mut names: Vec<String>) -> Result<Arc<[String]>, ()> {
+    if names.iter().any(|name| {
+        name.is_empty()
+            || name.len() > MAX_TIER_REGISTRY_NAME_BYTES
+            || name.bytes().any(|byte| byte.is_ascii_control())
+            || name == UNKNOWN_TIER
+            || name == storageclass::STANDARD
+            || name == storageclass::RRS
+    }) {
+        return Err(());
+    }
+    names.sort_unstable();
+    if names.windows(2).any(|pair| pair[0] == pair[1]) {
+        return Err(());
+    }
+    Ok(names.into())
+}
 
 /// Tier names currently registered in the tier configuration, cached for
 /// `TIER_NAME_CACHE_TTL`.
 pub(crate) async fn runtime_tier_names() -> Arc<[String]> {
+    runtime_tier_registry().await.names
+}
+
+/// Return the immutable tier registry for one scanner cycle/leader pair.
+/// Different buckets and disks belonging to the same cycle share this entry,
+/// so a TTL refresh cannot split one published cycle across generations.
+pub(crate) async fn runtime_tier_registry_for_cycle(cycle: u64, leader_epoch: u64) -> TierRegistrySnapshot {
+    let key = (cycle, leader_epoch);
+    prune_inactive_tier_cycle_snapshots(cycle, leader_epoch);
     {
-        let cached = TIER_NAME_CACHE.read().unwrap_or_else(|err| err.into_inner()).clone();
-        if let Some((refreshed_at, names)) = cached
-            && refreshed_at.elapsed() < TIER_NAME_CACHE_TTL
-        {
-            return names;
+        let cached = TIER_CYCLE_SNAPSHOTS.read().unwrap_or_else(|err| err.into_inner());
+        if let Some(snapshot) = cached.get(&key) {
+            return snapshot.clone();
         }
     }
 
-    let tiers = ecstore_get_global_tier_config_mgr().read().await.list_tiers();
-    let names: Arc<[String]> = tiers.iter().map(|tier| tier.name.clone()).collect::<Vec<_>>().into();
-    *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((Instant::now(), Arc::clone(&names)));
-    names
+    let mut snapshot = runtime_tier_registry().await;
+    // The registry generation describes the configuration snapshot, not the
+    // scan that consumed it. Keep it stable across cycles so a healthy cache
+    // can be reused; cycle and leader fencing are carried separately by the
+    // cache metadata and scan plan.
+    snapshot.generation = tier_registry_generation(&snapshot.names);
+    let mut cached = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    if let Some(existing) = cached.get(&key) {
+        return existing.clone();
+    }
+    cached.insert(key, snapshot.clone());
+    snapshot
+}
+
+fn prune_inactive_tier_cycle_snapshots(cycle: u64, leader_epoch: u64) {
+    let active = TIER_ACTIVE_CYCLES.read().unwrap_or_else(|err| err.into_inner());
+    let mut snapshots = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    snapshots.retain(|(entry_cycle, entry_epoch), _| {
+        active.contains_key(&(*entry_cycle, *entry_epoch))
+            || *entry_epoch > leader_epoch
+            || (*entry_epoch == leader_epoch && *entry_cycle >= cycle)
+    });
+}
+
+pub(crate) struct TierRegistryCycleGuard {
+    key: (u64, u64),
+}
+
+impl Drop for TierRegistryCycleGuard {
+    fn drop(&mut self) {
+        let mut active = TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner());
+        if let Some(count) = active.get_mut(&self.key) {
+            *count = count.saturating_sub(1);
+            if *count == 0 {
+                active.remove(&self.key);
+            }
+        }
+    }
+}
+
+pub(crate) fn begin_tier_registry_cycle(cycle: u64, leader_epoch: u64) -> TierRegistryCycleGuard {
+    let mut active = TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner());
+    let count = active.entry((cycle, leader_epoch)).or_default();
+    *count = count.saturating_add(1);
+    TierRegistryCycleGuard {
+        key: (cycle, leader_epoch),
+    }
+}
+
+fn tier_registry_generation(names: &[String]) -> u64 {
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-tier-registry-v1");
+    for name in names {
+        hasher.update(u64::try_from(name.len()).unwrap_or(u64::MAX).to_le_bytes());
+        hasher.update(name.as_bytes());
+    }
+    let digest = hasher.finalize();
+    let mut prefix = [0_u8; 8];
+    prefix.copy_from_slice(&digest[..8]);
+    u64::from_le_bytes(prefix)
+}
+
+/// Drop cycle snapshots only after the scanner has finished publishing a
+/// cycle. In-flight or retryable cycles must retain their original registry;
+/// TTL/capacity eviction could make a later bucket in the same cycle refresh
+/// to a different generation.
+pub(crate) fn complete_tier_registry_cycle(cycle: u64, leader_epoch: u64) {
+    let active = TIER_ACTIVE_CYCLES.read().unwrap_or_else(|err| err.into_inner());
+    let mut cached = TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner());
+    cached.retain(|(entry_cycle, entry_epoch), _| {
+        active.contains_key(&(*entry_cycle, *entry_epoch))
+            || *entry_epoch > leader_epoch
+            || (*entry_epoch == leader_epoch && *entry_cycle > cycle)
+    });
 }
 
 /// Test-only cache reset; the production cache has no invalidation hook
@@ -408,6 +619,9 @@ pub(crate) async fn runtime_tier_names() -> Arc<[String]> {
 #[cfg(test)]
 fn reset_tier_name_cache_for_test() {
     *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = None;
+    TIER_ACTIVE_CYCLES.write().unwrap_or_else(|err| err.into_inner()).clear();
+    TIER_CYCLE_SNAPSHOTS.write().unwrap_or_else(|err| err.into_inner()).clear();
+    TIER_REGISTRY_GENERATION.store(0, Ordering::Relaxed);
 }
 
 pub(crate) async fn enqueue_runtime_free_version(oi: ScannerObjectInfo) {
@@ -616,6 +830,68 @@ mod tests {
         assert!(Arc::ptr_eq(&first, &second));
     }
 
+    #[tokio::test]
+    async fn tier_registry_cycle_snapshot_stays_fixed_while_active() {
+        reset_tier_name_cache_for_test();
+        let cycle = 9_000_001;
+        let leader_epoch = 9_000_002;
+        let guard = begin_tier_registry_cycle(cycle, leader_epoch);
+        let first = runtime_tier_registry_for_cycle(cycle, leader_epoch).await;
+
+        // Simulate a TTL refresh observing a different configuration while the
+        // original cycle is still scanning. The active cycle entry must win.
+        *TIER_NAME_CACHE.write().unwrap_or_else(|err| err.into_inner()) = Some((
+            Instant::now() - TIER_NAME_CACHE_TTL - Duration::from_secs(1),
+            TierRegistrySnapshot {
+                generation: u64::MAX,
+                names: Arc::from(["COLD".to_string()]),
+                refresh_failed: false,
+            },
+        ));
+        let second = runtime_tier_registry_for_cycle(cycle, leader_epoch).await;
+        assert_eq!(second.generation, first.generation);
+        assert_eq!(second.names, first.names);
+
+        drop(guard);
+        complete_tier_registry_cycle(cycle, leader_epoch);
+        reset_tier_name_cache_for_test();
+    }
+
+    #[tokio::test]
+    async fn tier_registry_generation_survives_new_cycle_with_same_names() {
+        reset_tier_name_cache_for_test();
+        let first_cycle = 9_000_011;
+        let second_cycle = first_cycle + 1;
+        let leader_epoch = 9_000_012;
+
+        let first_guard = begin_tier_registry_cycle(first_cycle, leader_epoch);
+        let first = runtime_tier_registry_for_cycle(first_cycle, leader_epoch).await;
+        drop(first_guard);
+        complete_tier_registry_cycle(first_cycle, leader_epoch);
+
+        let second_guard = begin_tier_registry_cycle(second_cycle, leader_epoch);
+        let second = runtime_tier_registry_for_cycle(second_cycle, leader_epoch).await;
+        assert_eq!(first.names, second.names);
+        assert_eq!(first.generation, second.generation);
+
+        drop(second_guard);
+        complete_tier_registry_cycle(second_cycle, leader_epoch);
+        reset_tier_name_cache_for_test();
+    }
+
+    #[test]
+    fn invalid_tier_registry_names_fail_closed_for_refresh() {
+        assert!(validate_tier_registry_names(vec!["COLD\n".to_string()]).is_err());
+        assert!(validate_tier_registry_names(vec![UNKNOWN_TIER.to_string()]).is_err());
+        assert!(validate_tier_registry_names(vec!["COLD".to_string(), "COLD".to_string()]).is_err());
+        assert_eq!(
+            validate_tier_registry_names(vec!["WARM".to_string(), "COLD".to_string()])
+                .expect("valid registry names")
+                .as_ref(),
+            ["COLD".to_string(), "WARM".to_string()]
+        );
+    }
+
     #[test]
     fn foreground_read_guard_tracks_stream_lifetime() {
         reset_foreground_read_activity_for_test();

crates/scanner/src/remote_scanner/stream.rs

@@ -16,8 +16,8 @@
 use crate::RUSTFS_META_BUCKET;
 use crate::scanner_budget::{ScannerCycleBudget, ScannerCycleBudgetConfig};
 use crate::scanner_io::{
-    DataUsageCacheScanState, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks, cache_root_entry_info,
-    current_cache_root_or_prepare, scanner_set_disk_inventory,
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks,
+    cache_root_entry_info, current_cache_root_or_prepare_with_generation, scanner_set_disk_inventory,
 };
 use crate::storage_api::owner::NS_SCANNER_PROTOCOL_VERSION;
 use crate::{
@@ -212,6 +212,7 @@ pub(crate) struct RemoteScannerScanSpec<'a> {
     pub(crate) session_id: Uuid,
     pub(crate) session_sequence: u64,
     pub(crate) scan_plan_digest: DataUsageScanPlanDigest,
+    pub(crate) tier_registry_generation: u64,
     pub(crate) skip_healing: bool,
     pub(crate) scan_mode: HealScanMode,
 }
@@ -222,6 +223,7 @@ struct RemoteScannerResponseExpectation<'a> {
     source: DataUsageCacheSource,
     next_cycle: u64,
     scan_plan_digest: DataUsageScanPlanDigest,
+    tier_registry_generation: u64,
 }
 
 #[derive(Debug)]
@@ -668,6 +670,11 @@ async fn scan_and_persist_local_bucket(
         scan_mode,
         ..
     } = request;
+    // Keep the worker's cycle snapshot alive through cache reuse, scanning,
+    // and persistence. Without the guard, a later cycle can prune this key
+    // while this request is still running and allow a second registry to be
+    // selected for the same cycle.
+    let _tier_cycle_guard = crate::begin_tier_registry_cycle(next_cycle, leader_epoch);
     let store = resolve_scanner_object_store_handle()
         .ok_or_else(|| RemoteScannerServerError::worker("remote namespace scanner object layer is unavailable"))?;
     validate_remote_scanner_request_fence_with_store(next_cycle, leader_epoch, store.clone())
@@ -700,7 +707,25 @@ async fn scan_and_persist_local_bucket(
     let revisions = cache.load_with_revisions(set.clone(), &cache_name).await.map_err(|err| {
         RemoteScannerServerError::worker(format!("remote namespace scanner cache load or revision lookup failed: {err}"))
     })?;
-    let scan_state = current_cache_root_or_prepare(&mut cache, &bucket, source, next_cycle, leader_epoch, scan_plan_digest, true);
+    // Remote workers use the same cycle-frozen registry as `scan_data_folder`.
+    // Requiring its generation here prevents a cache snapshot classified by an
+    // older registry from being reused before the folder scan gets a chance to
+    // refresh it.
+    let tier_registry_generation = crate::runtime_tier_registry_for_cycle(next_cycle, leader_epoch)
+        .await
+        .generation;
+    let scan_state = current_cache_root_or_prepare_with_generation(
+        &mut cache,
+        &bucket,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        DataUsageCacheReuseOptions {
+            require_source: true,
+            tier_registry_generation: Some(tier_registry_generation),
+        },
+    );
     match scan_state {
         DataUsageCacheScanState::Current(usage) => {
             if guard.is_lock_lost() {
@@ -844,6 +869,7 @@ pub(crate) async fn scan_remote_bucket(
         session_id,
         session_sequence,
         scan_plan_digest,
+        tier_registry_generation,
         skip_healing,
         scan_mode,
     } = spec;
@@ -932,6 +958,7 @@ pub(crate) async fn scan_remote_bucket(
             source: expected_source,
             next_cycle,
             scan_plan_digest,
+            tier_registry_generation,
         },
         authenticator,
         rpc_deadline,
@@ -987,6 +1014,7 @@ where
             source: expected_source,
             next_cycle: TEST_NEXT_CYCLE,
             scan_plan_digest: expected_scan_plan_digest,
+            tier_registry_generation: 0,
         },
         authenticator,
         Instant::now() + NS_SCANNER_MAX_RPC_LIFETIME,
@@ -1086,6 +1114,11 @@ where
                         "remote namespace scanner returned usage for a different bucket plan",
                     )));
                 }
+                if complete.usage.tier_registry_generation != Some(expected.tier_registry_generation) {
+                    return Err(RemoteScannerStreamError::reconciled(StorageError::other(
+                        "remote namespace scanner returned usage for a different tier registry generation",
+                    )));
+                }
                 if !complete.usage.entry.children.is_empty() {
                     return Err(RemoteScannerStreamError::reconciled(StorageError::other(
                         "remote namespace scanner returned non-flattened bucket usage",

crates/scanner/src/remote_scanner/stream/tests.rs

@@ -195,6 +195,7 @@ fn test_usage(bucket: &str, objects: usize) -> DataUsageEntryInfo {
         name: bucket.to_string(),
         parent: crate::DATA_USAGE_ROOT.to_string(),
         entry,
+        tier_registry_generation: Some(0),
     }
 }
 
@@ -751,6 +752,43 @@ async fn complete_terminal_frame_reconciles_progress_and_usage() {
     assert_eq!(budget.progress(), (3, 2));
 }
 
+#[tokio::test]
+async fn terminal_usage_from_a_different_tier_generation_is_rejected() {
+    let request_id = Uuid::new_v4();
+    let writer_auth = FrameAuthenticator::for_test(request_id);
+    let reader_auth = FrameAuthenticator::for_test(request_id);
+    let (mut writer, reader) = tokio::io::duplex(4096);
+    tokio::spawn(async move {
+        let mut usage = test_usage("bucket", 1);
+        usage.tier_registry_generation = Some(1);
+        let mut sequence = 0;
+        write_frame(
+            &mut writer,
+            &writer_auth,
+            &mut sequence,
+            &RemoteScannerFrame::terminal(
+                RemoteScannerProgress::default(),
+                RemoteScannerFrameResult::Complete(Box::new(RemoteScannerComplete {
+                    source: TEST_SOURCE,
+                    scan_plan_digest: TEST_PLAN_DIGEST,
+                    usage,
+                    pending_maintenance_work: false,
+                })),
+            ),
+        )
+        .await
+        .expect("terminal frame should write");
+    });
+
+    let parent = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&parent, ScannerCycleBudgetConfig::default());
+    let error = consume_remote_scanner_stream(reader, parent, budget, "bucket", TEST_SOURCE, TEST_PLAN_DIGEST, reader_auth)
+        .await
+        .expect_err("generation mismatch must fail closed");
+
+    assert!(error.to_string().contains("tier registry generation"));
+}
+
 #[tokio::test]
 async fn complete_terminal_frame_after_budget_expiry_is_partial() {
     let request_id = Uuid::new_v4();

crates/scanner/src/scanner_folder.rs

@@ -55,9 +55,9 @@ use tracing::{debug, error, warn};
 use crate::{
     Disk, DiskError, DiskInfoOptions, Evaluator, Event, LcEventSrc, ListPathRawOptions, ObjectOpts, ReplicationConfig,
     ReplicationHealObject, ReplicationQueueAdmission, ReplicationStatusType, STORAGE_FORMAT_FILE, ScannerDiskExt as _,
-    ScannerLifecycleConfigExt as _, ScannerVersioningConfigExt as _, StorageError, apply_expiry_rule, apply_transition_rule,
-    enqueue_runtime_newer_noncurrent, is_reserved_or_invalid_bucket, list_path_raw, path2_bucket_object,
-    path2_bucket_object_with_base_path, queue_replication_heal, scanner_is_erasure,
+    ScannerLifecycleConfigExt as _, ScannerVersioningConfigExt as _, StorageError, TierRegistrySnapshot, apply_expiry_rule,
+    apply_transition_rule, enqueue_runtime_newer_noncurrent, is_reserved_or_invalid_bucket, list_path_raw, path2_bucket_object,
+    path2_bucket_object_with_base_path, queue_replication_heal, runtime_tier_registry_for_cycle, scanner_is_erasure,
     scanner_replication_config_for_lifecycle_eval,
 };
 use crate::{ScannerObjectInfo as ObjectInfo, ScannerObjectToDelete as ObjectToDelete};
@@ -626,6 +626,20 @@ fn apply_scanner_size_summary(into: &mut DataUsageEntry, summary: &SizeSummary)
     }
 
     into.add_tier_sizes(&summary.tier_stats);
+    into.add_unknown_tier_stats(&summary.unknown_tier_stats);
+    into.tier_accounting_proof = match (into.tier_accounting_proof, Some(summary.tier_accounting_proof)) {
+        (Some(mut current), Some(next)) => {
+            current.saturating_add(next);
+            Some(current)
+        }
+        (Some(_), None) => None,
+        (None, next) => next,
+    };
+    if into.unknown_tier_stats.as_ref().is_some_and(|stats| stats.counter_overflowed)
+        && let Some(proof) = into.tier_accounting_proof.as_mut()
+    {
+        proof.overflowed = true;
+    }
 }
 
 fn data_usage_root_has_progress(root: &DataUsageEntry) -> bool {
@@ -636,6 +650,8 @@ fn data_usage_root_has_progress(root: &DataUsageEntry) -> bool {
         || root.delete_markers > 0
         || root.failed_objects > 0
         || root.replication_stats.is_some()
+        || root.all_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty())
+        || root.unknown_tier_stats.as_ref().is_some_and(|stats| !stats.is_empty())
 }
 
 fn partial_cache_is_useful(root: &DataUsageEntry, pending_heals_changed: bool) -> bool {
@@ -670,6 +686,9 @@ pub struct FolderScanner {
     budget: Arc<ScannerCycleBudget>,
     skip_heal: Arc<std::sync::atomic::AtomicBool>,
     local_disk: Arc<Disk>,
+    /// Tier registry frozen for this folder scan. A refresh applies to the
+    /// next scan and cannot mix generations in one aggregate.
+    tier_registry: TierRegistrySnapshot,
     pending_heals_changed: bool,
     #[cfg(test)]
     list_path_raw_options_observer: Option<mpsc::UnboundedSender<ListPathRawTimeoutSnapshot>>,
@@ -1329,7 +1348,11 @@ impl FolderScanner {
                     continue;
                 }
 
-                let sz = match self.local_disk.get_size(item.clone()).await {
+                let sz = match self
+                    .local_disk
+                    .get_size_with_tier_names(item.clone(), &self.tier_registry.names)
+                    .await
+                {
                     Ok(sz) => sz,
                     Err(e) => {
                         let failure_action = classify_get_size_failure(&item, &e);
@@ -2161,6 +2184,10 @@ pub async fn scan_data_folder(
 
     let failed_object_ttl = rustfs_utils::get_env_u32(ENV_FAILED_OBJECT_TTL_SECS, DEFAULT_FAILED_OBJECT_TTL_SECS) as u64;
     let failed_objects_max = rustfs_utils::get_env_u32(ENV_FAILED_OBJECTS_MAX, DEFAULT_FAILED_OBJECTS_MAX) as usize;
+    let tier_registry = runtime_tier_registry_for_cycle(cache.info.next_cycle, cache.info.leader_epoch).await;
+    let mut cache = cache;
+    cache.fold_retired_tiers(&tier_registry.names);
+    cache.info.tier_registry_generation = Some(tier_registry.generation);
 
     // Create folder scanner
     let mut scanner = FolderScanner {
@@ -2189,6 +2216,7 @@ pub async fn scan_data_folder(
         budget: budget.clone(),
         skip_heal,
         local_disk,
+        tier_registry,
         pending_heals_changed: false,
         #[cfg(test)]
         list_path_raw_options_observer: None,

crates/scanner/src/scanner_folder/item_actions.rs

@@ -289,11 +289,62 @@ impl ScannerItem {
         item.object_path()
     }
 
+    fn effective_tier(oi: &ObjectInfo) -> &str {
+        if oi.transitioned_object.status == crate::TRANSITION_COMPLETE {
+            oi.transitioned_object.tier.as_str()
+        } else {
+            oi.storage_class.as_deref().unwrap_or(crate::storageclass::STANDARD)
+        }
+    }
+
+    fn tier_name_is_known(tier: &str, tier_names: &[String]) -> bool {
+        !tier.is_empty()
+            && tier != crate::data_usage_define::UNKNOWN_TIER
+            && (tier == crate::storageclass::STANDARD
+                || tier == crate::storageclass::RRS
+                || tier_names.iter().any(|name| name == tier))
+    }
+
+    pub(crate) fn tier_is_known(oi: &ObjectInfo, tier_names: &[String]) -> bool {
+        Self::tier_name_is_known(Self::effective_tier(oi), tier_names)
+    }
+
+    fn action_requires_known_tier(action: IlmAction) -> bool {
+        matches!(
+            action,
+            IlmAction::TransitionAction
+                | IlmAction::TransitionVersionAction
+                | IlmAction::DeleteAction
+                | IlmAction::DeleteVersionAction
+                | IlmAction::DeleteRestoredAction
+                | IlmAction::DeleteRestoredVersionAction
+                | IlmAction::DeleteAllVersionsAction
+                | IlmAction::DelMarkerDeleteAllVersionsAction
+        )
+    }
+
+    fn action_blocked_by_unknown_tier(
+        action: IlmAction,
+        oi: &ObjectInfo,
+        all_versions_known: bool,
+        tier_names: &[String],
+        target: &str,
+    ) -> bool {
+        if !Self::action_requires_known_tier(action) {
+            return false;
+        }
+        !Self::tier_is_known(oi, tier_names)
+            || (action.delete_all() && !all_versions_known)
+            || (matches!(action, IlmAction::TransitionAction | IlmAction::TransitionVersionAction)
+                && !Self::tier_name_is_known(target, tier_names))
+    }
+
     pub async fn apply_actions(
         &mut self,
         object_infos: Vec<ObjectInfo>,
         lock_retention: Option<Arc<ObjectLockConfiguration>>,
         versioning_config: VersioningConfiguration,
+        tier_names: &[String],
         size_summary: &mut SizeSummary,
     ) {
         let object_path = self.object_path();
@@ -402,6 +453,9 @@ impl ScannerItem {
         let mut noncurrent_accounting: Vec<PendingScannerAccounting<'_>> = Vec::new();
         let mut cumulative_size = 0;
         let mut remaining_versions = object_infos.len();
+        let all_versions_known = object_infos
+            .iter()
+            .all(|candidate| Self::tier_is_known(candidate, tier_names));
         'eventLoop: {
             for (i, event) in events.iter().enumerate() {
                 let oi = &object_infos[i];
@@ -425,6 +479,18 @@ impl ScannerItem {
                 let mut size = actual_size;
                 let mut account_now = true;
 
+                // A retired/unknown source tier may point at a remote object
+                // that cannot be safely deleted or transitioned. Lifecycle
+                // evaluation is still useful for accounting, but all
+                // side-effecting tier actions fail closed until the registry
+                // recognizes the source again.
+                if Self::action_blocked_by_unknown_tier(event.action, oi, all_versions_known, tier_names, &event.storage_class) {
+                    size = self.heal_actions(oi, actual_size, size_summary).await;
+                    size_summary.actions_accounting(oi, size, actual_size);
+                    cumulative_size += size;
+                    continue;
+                }
+
                 match event.action {
                     IlmAction::DeleteAllVersionsAction | IlmAction::DelMarkerDeleteAllVersionsAction => {
                         debug!(
@@ -929,4 +995,49 @@ mod tests {
         assert_eq!(item.object_name, "object");
         assert_eq!(item.object_path(), "object");
     }
+
+    #[test]
+    fn unknown_tier_never_triggers_transition() {
+        let object = ObjectInfo {
+            storage_class: Some("retired-tier".to_string()),
+            ..Default::default()
+        };
+        let tier_names = ["WARM".to_string()];
+        assert!(!ScannerItem::tier_is_known(&object, &tier_names));
+        assert!(ScannerItem::action_requires_known_tier(IlmAction::TransitionAction));
+        assert!(ScannerItem::action_requires_known_tier(IlmAction::DeleteVersionAction));
+        assert!(ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::TransitionAction,
+            &object,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+        assert!(!ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::NoneAction,
+            &object,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+
+        let known = ObjectInfo {
+            storage_class: Some(crate::storageclass::STANDARD.to_string()),
+            ..Default::default()
+        };
+        assert!(ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::DeleteAllVersionsAction,
+            &known,
+            false,
+            &tier_names,
+            "WARM"
+        ));
+        assert!(!ScannerItem::action_blocked_by_unknown_tier(
+            IlmAction::TransitionAction,
+            &known,
+            true,
+            &tier_names,
+            crate::storageclass::STANDARD
+        ));
+    }
 }

crates/scanner/src/scanner_folder/tests.rs

@@ -325,6 +325,11 @@ async fn build_test_scanner() -> (FolderScanner, std::path::PathBuf) {
         budget: ScannerCycleBudget::new(&CancellationToken::new(), Default::default()),
         skip_heal: Arc::new(AtomicBool::new(false)),
         local_disk: disk,
+        tier_registry: crate::TierRegistrySnapshot {
+            generation: 0,
+            names: Arc::new([]),
+            refresh_failed: false,
+        },
         pending_heals_changed: false,
         list_path_raw_options_observer: None,
     };

crates/scanner/src/scanner_io.rs

@@ -57,8 +57,9 @@ use crate::storage_api::scanner_io::{BucketInfo, BucketOptions};
 use crate::{
     BucketTargetSys, BucketVersioningSys, Disk, DiskError, ECStore, EcstoreError as Error, EcstoreResult as Result,
     RUSTFS_META_BUCKET, ReplicationConfig, STORAGE_FORMAT_FILE, ScannerDiskExt as _, ScannerLifecycleConfigExt as _,
-    ScannerReplicationConfigExt as _, ScannerVersioningConfigExt as _, SetDisks, StorageError, enqueue_runtime_free_version,
-    get_lifecycle_config, get_object_lock_config, get_replication_config, runtime_tier_names, storageclass,
+    ScannerReplicationConfigExt as _, ScannerVersioningConfigExt as _, SetDisks, StorageError, begin_tier_registry_cycle,
+    complete_tier_registry_cycle, enqueue_runtime_free_version, get_lifecycle_config, get_object_lock_config,
+    get_replication_config, runtime_tier_names, runtime_tier_registry_for_cycle, storageclass,
 };
 
 pub(crate) const SCANNER_SKIP_FILE_ERROR: &str = "skip file";
@@ -143,6 +144,7 @@ pub struct ScannerBucketScanPlan {
     all_buckets: Arc<Vec<BucketInfo>>,
     digest: DataUsageScanPlanDigest,
     leader_epoch: u64,
+    tier_registry_generation: u64,
     dirty_usage_buckets: Arc<DirtyUsageBuckets>,
     bucket_failures: ScannerBucketFailureState,
     pending_maintenance_work: Arc<AtomicBool>,
@@ -341,12 +343,20 @@ pub(crate) fn cache_root_entry_info(cache: &DataUsageCache) -> std::result::Resu
         name: cache.info.name.clone(),
         parent: DATA_USAGE_ROOT.to_string(),
         entry,
+        tier_registry_generation: cache.info.tier_registry_generation,
     })
 }
 
-fn apply_bucket_result_to_cache(cache: &mut DataUsageCache, result: DataUsageEntryInfo, update_time: SystemTime) {
+fn apply_bucket_result_to_cache(cache: &mut DataUsageCache, result: DataUsageEntryInfo, update_time: SystemTime) -> bool {
+    if cache.info.tier_registry_generation != result.tier_registry_generation {
+        // A result from another registry generation must never be folded into
+        // this cycle. Leaving it unapplied makes the cycle incomplete and
+        // forces the caller to re-account it under one frozen registry.
+        return false;
+    }
     cache.replace(&result.name, &result.parent, result.entry);
     cache.info.last_update = Some(update_time);
+    true
 }
 
 fn should_publish_completed_snapshot(completed_count: usize, total_count: usize, budget_elapsed: bool, cancelled: bool) -> bool {
@@ -498,6 +508,9 @@ pub trait ScannerIODisk: Send + Sync + Debug + 'static {
     ) -> Result<ScannerDiskScanOutcome>;
 
     async fn get_size(&self, item: ScannerItem) -> Result<SizeSummary>;
+
+    /// Read one object using a registry snapshot captured at scan start.
+    async fn get_size_with_tier_names(&self, item: ScannerItem, tier_names: &[String]) -> Result<SizeSummary>;
 }
 
 #[derive(Debug)]
@@ -649,7 +662,10 @@ use cache::*;
 use dirty_usage::*;
 use guards::*;
 
-pub(crate) use cache::{DataUsageCacheScanState, acquire_scanner_cache_locks, current_cache_root_or_prepare};
+pub(crate) use cache::{
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, acquire_scanner_cache_locks,
+    current_cache_root_or_prepare_with_generation,
+};
 pub use dirty_usage::{
     ScannerDirtyUsageAckError, ScannerDirtyUsageState, acknowledge_dirty_usage_generation, clear_dirty_usage_bucket,
     record_dirty_usage_bucket, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_state,

crates/scanner/src/scanner_io/cache.rs

@@ -100,13 +100,14 @@ where
     let _ = tokio::time::timeout(SCANNER_CACHE_LOCK_LOSS_SHUTDOWN_TIMEOUT, scan).await;
 }
 
-pub(crate) fn current_cache_root_entry(
+pub(crate) fn current_cache_root_entry_with_generation(
     cache: &DataUsageCache,
     name: &str,
     source: DataUsageCacheSource,
     next_cycle: u64,
     leader_epoch: u64,
     scan_plan_digest: DataUsageScanPlanDigest,
+    tier_registry_generation: Option<u64>,
 ) -> std::result::Result<Option<DataUsageEntryInfo>, ScannerError> {
     let metadata_is_current = cache.info.name == name
         && cache.info.source == Some(source)
@@ -115,7 +116,8 @@ pub(crate) fn current_cache_root_entry(
         && cache.info.last_update.is_some()
         && cache.info.next_cycle == next_cycle
         && cache.info.leader_epoch == leader_epoch
-        && cache.info.cache_key_format == DATA_USAGE_CACHE_KEY_FORMAT;
+        && cache.info.cache_key_format == DATA_USAGE_CACHE_KEY_FORMAT
+        && tier_registry_generation.is_none_or(|generation| cache.info.tier_registry_generation == Some(generation));
     if !metadata_is_current {
         return Ok(None);
     }
@@ -131,6 +133,13 @@ pub(crate) enum DataUsageCacheScanState {
     },
 }
 
+#[derive(Clone, Copy, Debug, Default)]
+pub(crate) struct DataUsageCacheReuseOptions {
+    pub(crate) require_source: bool,
+    pub(crate) tier_registry_generation: Option<u64>,
+}
+
+#[cfg(test)]
 pub(crate) fn current_cache_root_or_prepare(
     cache: &mut DataUsageCache,
     name: &str,
@@ -140,11 +149,51 @@ pub(crate) fn current_cache_root_or_prepare(
     scan_plan_digest: DataUsageScanPlanDigest,
     require_source: bool,
 ) -> DataUsageCacheScanState {
-    match current_cache_root_entry(cache, name, source, next_cycle, leader_epoch, scan_plan_digest) {
+    current_cache_root_or_prepare_with_generation(
+        cache,
+        name,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        DataUsageCacheReuseOptions {
+            require_source,
+            tier_registry_generation: None,
+        },
+    )
+}
+
+pub(crate) fn current_cache_root_or_prepare_with_generation(
+    cache: &mut DataUsageCache,
+    name: &str,
+    source: DataUsageCacheSource,
+    next_cycle: u64,
+    leader_epoch: u64,
+    scan_plan_digest: DataUsageScanPlanDigest,
+    options: DataUsageCacheReuseOptions,
+) -> DataUsageCacheScanState {
+    if options.tier_registry_generation.is_some_and(|generation| {
+        cache.info.next_cycle <= next_cycle
+            && cache.info.leader_epoch <= leader_epoch
+            && cache.info.tier_registry_generation != Some(generation)
+    }) {
+        // Make prepare_for_scan take its reset path so an entry classified by
+        // an older registry cannot be reused under the new cycle generation.
+        cache.info.scan_plan_digest = None;
+    }
+    match current_cache_root_entry_with_generation(
+        cache,
+        name,
+        source,
+        next_cycle,
+        leader_epoch,
+        scan_plan_digest,
+        options.tier_registry_generation,
+    ) {
         Ok(Some(root)) => DataUsageCacheScanState::Current(Box::new(root)),
         current => DataUsageCacheScanState::Prepared {
             invalid_current: current.err(),
-            outcome: cache.prepare_for_scan(name, next_cycle, leader_epoch, source, scan_plan_digest, require_source),
+            outcome: cache.prepare_for_scan(name, next_cycle, leader_epoch, source, scan_plan_digest, options.require_source),
         },
     }
 }
@@ -159,7 +208,7 @@ pub(super) fn cache_snapshot_is_current(
     scan_plan_digest: DataUsageScanPlanDigest,
 ) -> bool {
     matches!(
-        current_cache_root_entry(cache, name, source, next_cycle, leader_epoch, scan_plan_digest),
+        current_cache_root_entry_with_generation(cache, name, source, next_cycle, leader_epoch, scan_plan_digest, None),
         Ok(Some(_))
     )
 }
@@ -168,6 +217,7 @@ pub(super) fn completed_data_usage_info(
     results: &[DataUsageCache],
     expected_sources: &HashSet<DataUsageCacheSource>,
     all_buckets: &[String],
+    tier_registry_names: &[String],
     bucket_plan_complete: bool,
     budget_elapsed: bool,
     cancelled: bool,
@@ -183,6 +233,19 @@ pub(super) fn completed_data_usage_info(
         return None;
     }
 
+    // A generation is comparable across nodes because it is derived from the
+    // frozen registry names. Cycle and leader fencing remain separate cache
+    // metadata. Legacy peers omit the generation; an all-legacy result remains
+    // readable, but mixing legacy and new (or two new generations) would make
+    // the per-tier accounting ambiguous.
+    let registry_generation = results.first()?.info.tier_registry_generation;
+    if results.iter().any(|result| match registry_generation {
+        Some(generation) => result.info.tier_registry_generation != Some(generation),
+        None => result.info.tier_registry_generation.is_some(),
+    }) {
+        return None;
+    }
+
     if results.iter().any(|result| result.root().is_none()) {
         return None;
     }
@@ -200,9 +263,16 @@ pub(super) fn completed_data_usage_info(
         if !total.checked_merge(&merged) {
             return None;
         }
+        if !tier_accounting_proof_is_publishable(&merged, registry_generation, tier_registry_names) {
+            return None;
+        }
         buckets_usage.insert(bucket.clone(), checked_bucket_usage_info(&merged)?);
     }
 
+    if !tier_accounting_proof_is_publishable(&total, registry_generation, tier_registry_names) {
+        return None;
+    }
+
     let merged_last_update = results.iter().filter_map(|result| result.info.last_update).max()?;
     let bucket_sizes = buckets_usage
         .iter()
@@ -216,6 +286,7 @@ pub(super) fn completed_data_usage_info(
         delete_markers_total_count: u64::try_from(total.delete_markers).ok()?,
         objects_total_size: u64::try_from(total.size).ok()?,
         tier_stats: total.all_tier_stats.filter(|tiers| !tiers.is_empty()),
+        unknown_tier_stats: total.unknown_tier_stats.filter(|stats| !stats.is_empty()),
         buckets_count: u64::try_from(all_buckets.len()).ok()?,
         bucket_sizes,
         buckets_usage,
@@ -225,6 +296,109 @@ pub(super) fn completed_data_usage_info(
     Some((data_usage_info, merged_last_update))
 }
 
+fn tier_accounting_proof_is_publishable(
+    entry: &DataUsageEntry,
+    registry_generation: Option<u64>,
+    tier_registry_names: &[String],
+) -> bool {
+    let has_scalar_usage = entry.size > 0
+        || entry.objects > 0
+        || entry.versions > 0
+        || entry.delete_markers > 0
+        || entry.failed_objects > 0
+        || !entry.obj_sizes.is_empty()
+        || !entry.obj_versions.is_empty()
+        || entry.replication_stats.as_ref().is_some_and(|stats| !stats.is_empty());
+    let has_tier_accounted_data = entry
+        .all_tier_stats
+        .as_ref()
+        .is_some_and(|stats| stats.tiers.values().any(|tier| !tier.is_empty()))
+        || entry.unknown_tier_stats.as_ref().is_some_and(|stats| {
+            stats.counter_overflowed
+                || stats.unknown_bytes > 0
+                || stats.unknown_physical_bytes > 0
+                || stats.unknown_objects > 0
+                || stats.unknown_versions > 0
+        });
+
+    let Some(proof) = entry.tier_accounting_proof else {
+        return !has_scalar_usage && !has_tier_accounted_data;
+    };
+    if proof.overflowed
+        || entry
+            .unknown_tier_stats
+            .as_ref()
+            .is_some_and(|stats| stats.counter_overflowed)
+        || u64::try_from(entry.size).ok() != Some(proof.logical_total)
+    {
+        return false;
+    }
+    let unknown_logical = entry.unknown_tier_stats.as_ref().map_or(0, |stats| stats.unknown_bytes);
+    let unknown_physical = entry
+        .unknown_tier_stats
+        .as_ref()
+        .map_or(0, |stats| stats.unknown_physical_bytes);
+    if proof
+        .logical_known
+        .checked_add(unknown_logical)
+        .is_none_or(|total| total != proof.logical_total)
+        || proof
+            .physical_known
+            .checked_add(unknown_physical)
+            .is_none_or(|total| total != proof.physical_total)
+    {
+        return false;
+    }
+    if registry_generation.is_some()
+        && entry.all_tier_stats.as_ref().is_some_and(|stats| {
+            stats.tiers.keys().any(|tier| {
+                tier != crate::UNKNOWN_TIER
+                    && tier != crate::storageclass::STANDARD
+                    && tier != crate::storageclass::RRS
+                    && !tier_registry_names.iter().any(|allowed| allowed == tier)
+            })
+        })
+    {
+        return false;
+    }
+
+    if !has_tier_accounted_data {
+        return true;
+    }
+
+    let Some(tiers) = entry.all_tier_stats.as_ref() else {
+        return false;
+    };
+    let map_unknown_physical = tiers.tiers.get(crate::UNKNOWN_TIER).map_or(0, |stats| stats.total_size);
+    let companion_unknown_physical = entry
+        .unknown_tier_stats
+        .as_ref()
+        .map_or(0, |stats| stats.unknown_physical_bytes);
+    if map_unknown_physical != companion_unknown_physical {
+        return false;
+    }
+
+    // A no-configuration scan intentionally stores only UNKNOWN_TIER after
+    // the first unknown object; STANDARD/RRS remain absent to preserve the
+    // historical empty-map shape. In that shape the scalar proof is the sole
+    // source of known physical bytes. Configured registries seed at least one
+    // non-UNKNOWN key, whose map total must match the proof.
+    let has_known_tier_map = tiers.tiers.keys().any(|tier| tier.as_str() != crate::UNKNOWN_TIER);
+    if !has_known_tier_map {
+        return true;
+    }
+    let Some(known_tier_physical_total) = tiers
+        .tiers
+        .iter()
+        .filter(|(tier, _)| tier.as_str() != crate::UNKNOWN_TIER)
+        .map(|(_, stats)| stats)
+        .try_fold(0_u64, |total, stats| total.checked_add(stats.total_size))
+    else {
+        return false;
+    };
+    proof.physical_known == known_tier_physical_total
+}
+
 pub(super) async fn send_cache_root_entry_info(
     bucket_result_tx: &mpsc::Sender<DataUsageEntryInfo>,
     cache: &DataUsageCache,
@@ -319,13 +493,14 @@ pub(super) async fn persist_and_publish_cache_snapshot(
         return None;
     }
     if matches!(
-        current_cache_root_entry(
+        current_cache_root_entry_with_generation(
             &persisted,
             DATA_USAGE_ROOT,
             source,
             cache_snapshot.info.next_cycle,
             cache_snapshot.info.leader_epoch,
             scan_plan_digest,
+            cache_snapshot.info.tier_registry_generation,
         ),
         Ok(Some(_))
     ) {

crates/scanner/src/scanner_io/io_cache.rs

@@ -31,6 +31,7 @@ impl ScannerIOCache for SetDisks {
             all_buckets,
             digest: scan_plan_digest,
             leader_epoch,
+            tier_registry_generation,
             dirty_usage_buckets,
             bucket_failures,
             pending_maintenance_work,
@@ -48,6 +49,7 @@ impl ScannerIOCache for SetDisks {
                     next_cycle: want_cycle,
                     last_update: Some(now),
                     leader_epoch,
+                    tier_registry_generation: Some(tier_registry_generation),
                     source: Some(source),
                     snapshot_complete: true,
                     scan_plan_digest: Some(scan_plan_digest),
@@ -218,6 +220,13 @@ impl ScannerIOCache for SetDisks {
                 "Scanner old data usage cache load failed; rebuilding from bucket caches"
             );
         }
+        // Fence a stale set aggregate before copying entries into per-bucket work caches.
+        if old_cache.info.next_cycle <= want_cycle
+            && old_cache.info.leader_epoch <= leader_epoch
+            && old_cache.info.tier_registry_generation != Some(tier_registry_generation)
+        {
+            old_cache.info.scan_plan_digest = None;
+        }
         match old_cache.prepare_for_scan(
             DATA_USAGE_ROOT,
             want_cycle,
@@ -267,6 +276,7 @@ impl ScannerIOCache for SetDisks {
                 name: DATA_USAGE_ROOT.to_string(),
                 next_cycle: want_cycle,
                 leader_epoch,
+                tier_registry_generation: Some(tier_registry_generation),
                 source: Some(source),
                 snapshot_complete: false,
                 scan_plan_digest: Some(scan_plan_digest),
@@ -328,8 +338,9 @@ impl ScannerIOCache for SetDisks {
                         };
 
                         let mut cache = cache_mutex_clone.lock().await;
-                        apply_bucket_result_to_cache(&mut cache, result, SystemTime::now());
-                        completed_bucket_count_clone.fetch_add(1, Ordering::Relaxed);
+                        if apply_bucket_result_to_cache(&mut cache, result, SystemTime::now()) {
+                            completed_bucket_count_clone.fetch_add(1, Ordering::Relaxed);
+                        }
                     }
                 }
             }
@@ -460,6 +471,7 @@ impl ScannerIOCache for SetDisks {
                                 session_id: remote_session_id,
                                 session_sequence: request_sequence,
                                 scan_plan_digest: bucket_scan_plan_digest,
+                                tier_registry_generation,
                                 skip_healing: healing,
                                 scan_mode,
                             },
@@ -675,14 +687,17 @@ impl ScannerIOCache for SetDisks {
                             continue;
                         }
                     };
-                    let scan_state = current_cache_root_or_prepare(
+                    let scan_state = current_cache_root_or_prepare_with_generation(
                         &mut cache,
                         &bucket.name,
                         source,
                         want_cycle,
                         leader_epoch,
                         bucket_scan_plan_digest,
-                        require_cache_source,
+                        DataUsageCacheReuseOptions {
+                            require_source: require_cache_source,
+                            tier_registry_generation: Some(tier_registry_generation),
+                        },
                     );
                     let outcome = match scan_state {
                         DataUsageCacheScanState::Current(root) => {
@@ -1108,6 +1123,7 @@ impl ScannerIOCache for SetDisks {
                     name: DATA_USAGE_ROOT.to_string(),
                     next_cycle: want_cycle,
                     leader_epoch,
+                    tier_registry_generation: Some(tier_registry_generation),
                     source: Some(source),
                     snapshot_complete: false,
                     scan_plan_digest: Some(scan_plan_digest),

crates/scanner/src/scanner_io/io_cycle.rs

@@ -48,6 +48,7 @@ impl ScannerIOCycle for ECStore {
         scan_mode: HealScanMode,
     ) -> Result<ScannerCycleResult> {
         let child_token = ctx.child_token();
+        let _tier_cycle_guard = begin_tier_registry_cycle(want_cycle, leader_epoch);
 
         // Check the local pool metadata before listing buckets. A failed or
         // canceled decommission remains suspended after its worker exits, so
@@ -127,6 +128,8 @@ impl ScannerIOCycle for ECStore {
             scanner_bucket_plan_digest(&all_buckets, crate::scanner::scanner_activity_snapshot_digest(&activity_before));
         let dirty_usage_snapshot = Arc::new(snapshot_dirty_usage_buckets(&all_buckets, dirty_generation_before_bucket_list));
         let cache_cycle_floor = Arc::new(AtomicU64::new(want_cycle));
+        let tier_registry = runtime_tier_registry_for_cycle(want_cycle, leader_epoch).await;
+        let tier_registry_generation = tier_registry.generation;
 
         if all_buckets.is_empty() {
             reset_set_scan_gauges();
@@ -157,6 +160,9 @@ impl ScannerIOCycle for ECStore {
             {
                 return Ok(ScannerCycleResult::new(status, None));
             }
+            if status == ScannerCycleStatus::Complete {
+                complete_tier_registry_cycle(want_cycle, leader_epoch);
+            }
             let dirty_usage_clear =
                 (status == ScannerCycleStatus::Complete).then(|| dirty_usage_snapshot.buckets.as_ref().clone());
             let remote_dirty_usage_acknowledgements = if status == ScannerCycleStatus::Complete {
@@ -249,6 +255,7 @@ impl ScannerIOCycle for ECStore {
                     all_buckets: Arc::clone(&all_buckets),
                     digest: scan_plan_digest,
                     leader_epoch,
+                    tier_registry_generation,
                     dirty_usage_buckets: dirty_usage_snapshot.buckets.clone(),
                     bucket_failures: bucket_failures.clone(),
                     pending_maintenance_work: pending_maintenance_work.clone(),
@@ -366,6 +373,7 @@ impl ScannerIOCycle for ECStore {
             &results,
             &expected_sources,
             &all_bucket_names,
+            &tier_registry.names,
             bucket_plan_complete,
             budget_elapsed,
             ctx.is_cancelled(),
@@ -391,6 +399,9 @@ impl ScannerIOCycle for ECStore {
             &failed_buckets,
         );
         result?;
+        if cycle_status == ScannerCycleStatus::Complete {
+            complete_tier_registry_cycle(want_cycle, leader_epoch);
+        }
         let remote_dirty_usage_acknowledgements = if cycle_status == ScannerCycleStatus::Complete {
             crate::scanner::scanner_dirty_usage_acknowledgements(&activity_before)
         } else {

crates/scanner/src/scanner_io/io_disk.rs

@@ -13,29 +13,38 @@
 // limitations under the License.
 /// ScannerIODisk implementation for Disk: get_size and the per-disk bucket scan.
 use super::*;
+use crate::UNKNOWN_TIER;
 
 ///
 /// Seed [`SizeSummary::tier_stats`] from the cached tier-name list.
 ///
-/// Preserves the original seeding semantics: with no tiers configured the map
-/// stays completely empty (STANDARD/RRS are not seeded either); otherwise the
-/// standard storage classes are seeded alongside every configured tier so
-/// per-object accounting always finds its tier key.
+/// Preserves the original no-tier shape: with no tiers configured the map
+/// stays completely empty (STANDARD/RRS/UNKNOWN are not seeded either).
+/// Otherwise the standard storage classes and one fixed unknown bucket are
+/// seeded alongside every configured tier so per-object accounting never
+/// inserts an untrusted metadata key.
 pub(super) fn tier_stats_template(tier_names: &[String]) -> HashMap<String, TierStats> {
-    let mut tier_stats = HashMap::with_capacity(tier_names.len() + 2);
+    let mut tier_stats = HashMap::with_capacity(tier_names.len() + 3);
     for tier_name in tier_names {
-        tier_stats.insert(tier_name.clone(), TierStats::default());
+        if tier_name != UNKNOWN_TIER {
+            tier_stats.insert(tier_name.clone(), TierStats::default());
+        }
     }
     if !tier_stats.is_empty() {
         tier_stats.insert(storageclass::STANDARD.to_string(), TierStats::default());
         tier_stats.insert(storageclass::RRS.to_string(), TierStats::default());
+        tier_stats.insert(UNKNOWN_TIER.to_string(), TierStats::default());
     }
     tier_stats
 }
 
 #[async_trait::async_trait]
 impl ScannerIODisk for Disk {
-    async fn get_size(&self, mut item: ScannerItem) -> Result<SizeSummary> {
+    async fn get_size(&self, item: ScannerItem) -> Result<SizeSummary> {
+        self.get_size_with_tier_names(item, &runtime_tier_names().await).await
+    }
+
+    async fn get_size_with_tier_names(&self, mut item: ScannerItem, tier_names: &[String]) -> Result<SizeSummary> {
         let done_object = Metrics::time(Metric::ScanObject);
 
         if !is_xl_meta_path(&item.path) {
@@ -105,12 +114,13 @@ impl ScannerIODisk for Disk {
             .map(|v| ObjectInfo::from_file_info(v, item.bucket.as_str(), object_path.as_str(), versioned))
             .collect::<Vec<ObjectInfo>>();
 
-        let mut size_summary = SizeSummary::default();
-
-        // Tier names come from the process-wide TTL cache; seeding from them
-        // replaces the per-object clone of every full TierConfig.
-        let tier_names = runtime_tier_names().await;
-        size_summary.tier_stats = tier_stats_template(&tier_names);
+        // The caller supplies one registry snapshot for the whole folder scan;
+        // seeding from it prevents a TTL refresh from mixing generations in a
+        // single result.
+        let mut size_summary = SizeSummary {
+            tier_stats: tier_stats_template(tier_names),
+            ..Default::default()
+        };
 
         let lock_config = object_lock_config_for_scanner_item(&item).await;
 
@@ -120,12 +130,14 @@ impl ScannerIODisk for Disk {
         // `object_infos`.
         global_metrics().record_scanner_versions_scanned(object_infos.len() as u64);
 
-        item.apply_actions(object_infos, lock_config, versioning_config, &mut size_summary)
+        item.apply_actions(object_infos, lock_config, versioning_config, tier_names, &mut size_summary)
             .await;
 
         if !free_version_infos.is_empty() {
             for oi in free_version_infos {
-                enqueue_runtime_free_version(oi).await;
+                if ScannerItem::tier_is_known(&oi, tier_names) {
+                    enqueue_runtime_free_version(oi).await;
+                }
             }
         }
 

crates/scanner/src/scanner_io/publish_gate_tests.rs

@@ -13,7 +13,8 @@
 // limitations under the License.
 
 use super::*;
-use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage};
+use crate::data_usage_define::{UNKNOWN_TIER, UnknownTierStats, hash_path};
+use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage, TierAccountingProof};
 
 const TEST_PLAN_DIGEST: DataUsageScanPlanDigest = DataUsageScanPlanDigest([7; 32]);
 
@@ -89,6 +90,11 @@ fn completed_root_cache(bucket: &str, objects: usize, update_secs: u64, source:
         DataUsageEntry {
             objects,
             size: objects.saturating_mul(10),
+            tier_accounting_proof: Some(TierAccountingProof {
+                logical_total: u64::try_from(objects.saturating_mul(10)).unwrap_or(u64::MAX),
+                logical_known: u64::try_from(objects.saturating_mul(10)).unwrap_or(u64::MAX),
+                ..Default::default()
+            }),
             ..Default::default()
         },
     );
@@ -102,7 +108,7 @@ fn completed_data_usage_info_for_test(
     cancelled: bool,
 ) -> Option<(DataUsageInfo, SystemTime)> {
     let expected_sources = results.iter().filter_map(|result| result.info.source).collect::<HashSet<_>>();
-    completed_data_usage_info(results, &expected_sources, all_buckets, true, budget_elapsed, cancelled)
+    completed_data_usage_info(results, &expected_sources, all_buckets, &[], true, budget_elapsed, cancelled)
 }
 
 #[test]
@@ -122,11 +128,21 @@ fn completed_data_usage_info_publishes_tier_stats_across_sets() {
     let mut first_set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
     let mut tiered = DataUsageEntry::default();
     tiered.add_tier_sizes(&warm(100, 2, 1));
+    tiered.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 100,
+        physical_known: 100,
+        ..Default::default()
+    });
     first_set.replace("bucket-b", DATA_USAGE_ROOT, tiered);
 
     let mut second_set = completed_root_cache("bucket-b", 2, 20, DataUsageCacheSource::new(1, 0));
     let mut tiered = DataUsageEntry::default();
     tiered.add_tier_sizes(&warm(50, 1, 1));
+    tiered.tier_accounting_proof = Some(TierAccountingProof {
+        physical_total: 50,
+        physical_known: 50,
+        ..Default::default()
+    });
     second_set.replace("bucket-a", DATA_USAGE_ROOT, tiered);
 
     let (data_usage_info, _) = completed_data_usage_info_for_test(&[first_set, second_set], &all_buckets, false, false)
@@ -145,6 +161,269 @@ fn completed_data_usage_info_publishes_tier_stats_across_sets() {
     );
 }
 
+#[test]
+fn completed_data_usage_info_rejects_logical_proof_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 9,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_logical_total_size_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 11;
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_physical_proof_mismatch() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 9,
+        physical_known: 9,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_unknown_physical_double_accounting() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_physical_bytes: 9,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_unknown_counter_overflow() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        counter_overflowed: true,
+        ..Default::default()
+    });
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_accepts_no_tier_standard_empty_map_with_proof() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+
+    let (info, _) = completed_data_usage_info_for_test(&[set], &all_buckets, false, false)
+        .expect("no-tier STANDARD/RRS usage does not require a tier map");
+    assert!(info.tier_stats.is_none());
+}
+
+#[test]
+fn completed_data_usage_info_accepts_no_tier_unknown_and_standard_shape() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 13;
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 3,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 9,
+        unknown_physical_bytes: 3,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 13,
+        logical_known: 4,
+        physical_total: 7,
+        physical_known: 4,
+        ..Default::default()
+    });
+
+    let (info, _) = completed_data_usage_info_for_test(&[set], &all_buckets, false, false)
+        .expect("no-tier STANDARD plus UNKNOWN should remain publishable");
+    assert_eq!(
+        info.tier_stats.expect("unknown bucket should be retained").tiers[UNKNOWN_TIER].total_size,
+        3
+    );
+}
+
+#[test]
+fn completed_data_usage_info_accepts_unknown_only_with_current_registry_generation() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    set.info.tier_registry_generation = Some(7);
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.size = 13;
+    entry.add_tier_sizes(&HashMap::from([(
+        UNKNOWN_TIER.to_string(),
+        TierStats {
+            total_size: 3,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.add_unknown_tier_stats(&UnknownTierStats {
+        unknown_bytes: 9,
+        unknown_physical_bytes: 3,
+        unknown_objects: 1,
+        unknown_versions: 1,
+        ..Default::default()
+    });
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 13,
+        logical_known: 4,
+        physical_total: 7,
+        physical_known: 4,
+        ..Default::default()
+    });
+
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
+    assert!(
+        completed_data_usage_info(&[set], &expected_sources, &all_buckets, &["WARM".to_string()], true, false, false,).is_some()
+    );
+}
+
+#[test]
+fn completed_data_usage_info_rejects_non_registry_tier_in_current_generation() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    set.info.tier_registry_generation = Some(7);
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([
+        (
+            "WARM".to_string(),
+            TierStats {
+                total_size: 4,
+                num_versions: 1,
+                num_objects: 1,
+            },
+        ),
+        (
+            "RETIRED".to_string(),
+            TierStats {
+                total_size: 6,
+                num_versions: 1,
+                num_objects: 1,
+            },
+        ),
+    ]));
+    entry.tier_accounting_proof = Some(TierAccountingProof {
+        logical_total: 10,
+        logical_known: 10,
+        physical_total: 10,
+        physical_known: 10,
+        ..Default::default()
+    });
+
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
+    assert!(
+        completed_data_usage_info(&[set], &expected_sources, &all_buckets, &["WARM".to_string()], true, false, false,).is_none()
+    );
+}
+
+#[test]
+fn completed_data_usage_info_rejects_legacy_proof_missing_when_tier_accounted() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.add_tier_sizes(&HashMap::from([(
+        "WARM".to_string(),
+        TierStats {
+            total_size: 10,
+            num_versions: 1,
+            num_objects: 1,
+        },
+    )]));
+    entry.tier_accounting_proof = None;
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
+#[test]
+fn completed_data_usage_info_rejects_legacy_proof_missing_for_scalar_usage() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let mut set = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let entry = set.cache.get_mut(&hash_path("bucket-a").key()).expect("bucket entry");
+    entry.tier_accounting_proof = None;
+
+    assert!(completed_data_usage_info_for_test(&[set], &all_buckets, false, false).is_none());
+}
+
 #[test]
 fn completed_data_usage_info_omits_tier_stats_without_tiered_objects() {
     let all_buckets = vec!["bucket-a".to_string()];
@@ -156,6 +435,51 @@ fn completed_data_usage_info_omits_tier_stats_without_tiered_objects() {
     assert!(data_usage_info.tier_stats.is_none());
 }
 
+#[test]
+fn completed_data_usage_info_rejects_legacy_and_new_tier_generations_mixed() {
+    let all_buckets = vec!["bucket-a".to_string()];
+    let legacy = completed_root_cache("bucket-a", 1, 10, DataUsageCacheSource::new(0, 0));
+    let mut current = completed_root_cache("bucket-a", 1, 20, DataUsageCacheSource::new(1, 0));
+    current.info.tier_registry_generation = Some(42);
+
+    assert!(
+        completed_data_usage_info_for_test(&[legacy, current], &all_buckets, false, false).is_none(),
+        "legacy and generation-tagged sets must not publish a mixed snapshot"
+    );
+}
+
+#[test]
+fn current_cache_root_with_new_tier_generation_resets_old_cache() {
+    let source = DataUsageCacheSource::new(0, 0);
+    let mut cache = completed_root_cache("bucket-a", 1, 10, source);
+    cache.info.tier_registry_generation = Some(1);
+
+    let state = current_cache_root_or_prepare_with_generation(
+        &mut cache,
+        DATA_USAGE_ROOT,
+        source,
+        0,
+        0,
+        TEST_PLAN_DIGEST,
+        DataUsageCacheReuseOptions {
+            require_source: false,
+            tier_registry_generation: Some(2),
+        },
+    );
+
+    assert!(matches!(
+        state,
+        DataUsageCacheScanState::Prepared {
+            outcome: DataUsageCachePrepareOutcome::Reset,
+            ..
+        }
+    ));
+    assert!(cache.cache.is_empty(), "old-generation entries must not be reused");
+    assert_eq!(cache.info.tier_registry_generation, None);
+    assert_eq!(cache.info.scan_plan_digest, Some(TEST_PLAN_DIGEST));
+    assert!(!cache.info.snapshot_complete);
+}
+
 #[test]
 fn completed_data_usage_info_requires_every_set_before_publish() {
     let all_buckets = vec!["bucket-a".to_string(), "bucket-b".to_string(), "bucket-empty".to_string()];
@@ -280,6 +604,13 @@ fn completed_data_usage_info_flattens_nested_bucket_entries() {
             replica_size: 2048,
             replica_count: 2,
         }),
+        tier_accounting_proof: Some(TierAccountingProof {
+            logical_total: 2048,
+            logical_known: 2048,
+            physical_total: 2048,
+            physical_known: 2048,
+            ..Default::default()
+        }),
         ..Default::default()
     };
     nested.obj_sizes.add(2048);
@@ -347,7 +678,8 @@ fn completed_data_usage_info_requires_exact_topology_sources() {
     let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0), DataUsageCacheSource::new(1, 0)]);
 
     assert!(
-        completed_data_usage_info(&[first_set, unexpected_set], &expected_sources, &all_buckets, true, false, false).is_none()
+        completed_data_usage_info(&[first_set, unexpected_set], &expected_sources, &all_buckets, &[], true, false, false)
+            .is_none()
     );
 }
 
@@ -357,7 +689,7 @@ fn completed_data_usage_info_rejects_incomplete_bucket_plan() {
     let set = completed_root_cache("bucket", 2, 10, DataUsageCacheSource::new(0, 0));
     let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0)]);
 
-    assert!(completed_data_usage_info(&[set], &expected_sources, &all_buckets, false, false, false).is_none());
+    assert!(completed_data_usage_info(&[set], &expected_sources, &all_buckets, &[], false, false, false).is_none());
 }
 
 #[test]

crates/scanner/src/scanner_io/tests.rs

@@ -23,7 +23,7 @@ use crate::storage_api::owner::{
 use crate::storage_api::scan::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions, ObjectIO as _};
 use crate::{
     DiskOption, ECStore, Endpoint, EndpointServerPools, Endpoints, InstanceContext, PoolEndpoints, ScannerObjectOptions,
-    ScannerPutObjReader, init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests,
+    ScannerPutObjReader, UNKNOWN_TIER, init_bucket_metadata_sys_for_scanner_tests, init_ecstore_config_for_scanner_tests,
     init_local_disks_with_instance_ctx, new_disk, path2_bucket_object_with_base_path,
 };
 use rustfs_filemeta::FileInfo;
@@ -986,8 +986,8 @@ fn is_xl_meta_path_accepts_forward_separator() {
 fn tier_stats_template_seeds_tiers_and_standard_classes() {
     let template = tier_stats_template(&["WARM".to_string(), "COLD".to_string()]);
 
-    assert_eq!(template.len(), 4);
-    for tier in ["WARM", "COLD", storageclass::STANDARD, storageclass::RRS] {
+    assert_eq!(template.len(), 5);
+    for tier in ["WARM", "COLD", storageclass::STANDARD, storageclass::RRS, UNKNOWN_TIER] {
         assert_eq!(template.get(tier), Some(&TierStats::default()), "missing seed for tier {tier}");
     }
 }
@@ -1328,7 +1328,7 @@ fn apply_bucket_result_to_cache_updates_bucket_entry() {
     );
 
     let update_time = SystemTime::now();
-    apply_bucket_result_to_cache(
+    assert!(apply_bucket_result_to_cache(
         &mut cache,
         DataUsageEntryInfo {
             name: "bucket".to_string(),
@@ -1338,12 +1338,51 @@ fn apply_bucket_result_to_cache_updates_bucket_entry() {
                 objects: 2,
                 ..Default::default()
             },
+            tier_registry_generation: None,
         },
         update_time,
-    );
+    ));
 
     assert_eq!(cache.info.last_update, Some(update_time));
     let entry = cache.find("bucket").expect("bucket entry should remain present");
     assert_eq!(entry.size, 10);
     assert_eq!(entry.objects, 2);
 }
+
+#[test]
+fn apply_bucket_result_to_cache_rejects_a_different_tier_generation() {
+    let mut cache = DataUsageCache {
+        info: DataUsageCacheInfo {
+            name: DATA_USAGE_ROOT.to_string(),
+            tier_registry_generation: Some(7),
+            ..Default::default()
+        },
+        ..Default::default()
+    };
+    cache.replace(
+        "bucket",
+        DATA_USAGE_ROOT,
+        DataUsageEntry {
+            size: 3,
+            ..Default::default()
+        },
+    );
+
+    let applied = apply_bucket_result_to_cache(
+        &mut cache,
+        DataUsageEntryInfo {
+            name: "bucket".to_string(),
+            parent: DATA_USAGE_ROOT.to_string(),
+            entry: DataUsageEntry {
+                size: 11,
+                ..Default::default()
+            },
+            tier_registry_generation: Some(8),
+        },
+        SystemTime::now(),
+    );
+
+    assert!(!applied);
+    assert_eq!(cache.find("bucket").map(|entry| entry.size), Some(3));
+    assert!(cache.info.last_update.is_none());
+}

crates/storage-api/src/lib.rs

@@ -46,6 +46,7 @@ pub const NS_SCANNER_SERVER_EPOCH_QUERY: &str = "ns_scanner_server_epoch";
 pub const NS_SCANNER_SESSION_ID_QUERY: &str = "ns_scanner_session_id";
 pub const NS_SCANNER_SESSION_SEQUENCE_QUERY: &str = "ns_scanner_session_sequence";
 pub const NS_SCANNER_PROTOCOL_VERSION_QUERY: &str = "ns_scanner_protocol";
+pub const NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY: &str = "ns_scanner_tier_registry_generation";
 pub const NS_SCANNER_PROTOCOL_VERSION: u16 = 3;
 pub const SCANNER_ACTIVITY_LEGACY_PROTOCOL_VERSION: u32 = 0;
 pub const SCANNER_ACTIVITY_PREVIOUS_PROTOCOL_VERSION: u32 = 5;
@@ -57,6 +58,8 @@ pub struct NsScannerCapabilityResponse {
     pub version: u16,
     pub server_epoch: uuid::Uuid,
     pub proof: Vec<u8>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub supports_tier_registry_generation: Option<bool>,
 }
 
 pub mod admin;

crates/test-utils/src/lib.rs

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

deny.toml

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

docs/testing/ci-gates.md

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

rustfs/Cargo.toml

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

rustfs/src/admin/handlers/site_replication.rs

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

rustfs/src/admin/site_replication_identity.rs

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

rustfs/src/admin/storage_api.rs

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

rustfs/src/allocator_reclaim.rs

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

rustfs/src/main.rs

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

rustfs/src/memory_observability.rs

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

rustfs/src/storage/rpc/http_service.rs

@@ -19,13 +19,15 @@ use crate::storage::storage_api::rpc_consumer::http_service::{
     DEFAULT_READ_BUFFER_SIZE, DeleteOptions, DiskStore, NS_SCANNER_PROTOCOL_VERSION, NsScannerCapabilityResponse,
     PUT_FILE_AUTH_TRAILER_LEN, PUT_FILE_AUTH_V1, PUT_FILE_CAPABILITY_VERSION, PutFileCapabilityResponse, StorageDiskRpcExt as _,
     WALK_DIR_STREAM_COMPLETION_V1, WalkDirOptions, check_and_record_signed_rpc_nonce, find_local_disk_by_ref,
-    sign_ns_scanner_capability, sign_put_file_capability, verify_put_file_auth_trailer, verify_rpc_signature,
+    sign_ns_scanner_capability_with_tier_registry_generation, sign_put_file_capability, verify_put_file_auth_trailer,
+    verify_rpc_signature,
 };
 #[cfg(test)]
 use crate::storage::storage_api::rpc_consumer::http_service::{
     NS_SCANNER_BODY_SHA256_QUERY, NS_SCANNER_CAPABILITY_CHALLENGE_QUERY, NS_SCANNER_CYCLE_QUERY, NS_SCANNER_LEADER_EPOCH_QUERY,
     NS_SCANNER_REQUEST_ID_QUERY, NS_SCANNER_SERVER_EPOCH_QUERY, NS_SCANNER_SESSION_ID_QUERY, NS_SCANNER_SESSION_SEQUENCE_QUERY,
-    PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY, WALK_DIR_BODY_SHA256_QUERY,
+    NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY, PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY,
+    WALK_DIR_BODY_SHA256_QUERY,
 };
 use crate::storage::storage_api::runtime_sources_consumer::runtime_sources;
 use crate::storage::storage_api::tonic_rpc_auth_failure_reason;
@@ -290,6 +292,8 @@ struct NsScannerQuery {
 struct NsScannerCapabilityQuery {
     ns_scanner_protocol: Option<u16>,
     ns_scanner_challenge: Option<uuid::Uuid>,
+    #[serde(rename = "ns_scanner_tier_registry_generation")]
+    ns_scanner_tier_registry_generation: Option<bool>,
 }
 
 fn verify_ns_scanner_body_digest(query: &NsScannerQuery, body: &[u8]) -> bool {
@@ -410,7 +414,9 @@ async fn handle_internode_rpc(req: Request<Incoming>) -> Response<Body> {
         (Method::GET, WALK_DIR_PATH) | (Method::HEAD, WALK_DIR_PATH) => handle_walk_dir(req).await,
         (Method::GET, NS_SCANNER_PATH) => match parse_query::<NsScannerCapabilityQuery>(&req) {
             Ok(query) if query.ns_scanner_protocol == Some(NS_SCANNER_PROTOCOL_VERSION) => match query.ns_scanner_challenge {
-                Some(challenge) if !challenge.is_nil() => ns_scanner_capability_response(challenge),
+                Some(challenge) if !challenge.is_nil() => {
+                    ns_scanner_capability_response(challenge, query.ns_scanner_tier_registry_generation == Some(true))
+                }
                 Some(_) | None => response_with_status(StatusCode::BAD_REQUEST, "namespace scanner challenge is invalid"),
             },
             Ok(_) => response_with_status(StatusCode::UPGRADE_REQUIRED, "namespace scanner protocol is unsupported"),
@@ -466,31 +472,34 @@ fn record_internode_rpc_error(operation: Option<&'static str>) {
     }
 }
 
-fn ns_scanner_capability_response(challenge: uuid::Uuid) -> Response<Body> {
+fn ns_scanner_capability_response(challenge: uuid::Uuid, include_tier_registry_generation: bool) -> Response<Body> {
     let server_epoch = *NS_SCANNER_SERVER_EPOCH;
-    let proof = match sign_ns_scanner_capability(challenge, server_epoch) {
-        Ok(proof) => proof,
-        Err(err) => {
-            error!(
-                event = EVENT_RPC_REQUEST_FAILED,
-                component = LOG_COMPONENT_INTERNODE_RPC,
-                subsystem = LOG_SUBSYSTEM_NAMESPACE_SCANNER,
-                operation = INTERNODE_OPERATION_NS_SCANNER,
-                result = "failed",
-                status_code = StatusCode::UPGRADE_REQUIRED.as_u16(),
-                rpc_path = NS_SCANNER_PATH,
-                method = %Method::GET,
-                reason = "capability_authentication_unavailable",
-                error = %err,
-                "internode rpc request failed"
-            );
-            return response_with_status(StatusCode::UPGRADE_REQUIRED, "namespace scanner RPC authentication is unavailable");
-        }
-    };
+    let proof =
+        match sign_ns_scanner_capability_with_tier_registry_generation(challenge, server_epoch, include_tier_registry_generation)
+        {
+            Ok(proof) => proof,
+            Err(err) => {
+                error!(
+                    event = EVENT_RPC_REQUEST_FAILED,
+                    component = LOG_COMPONENT_INTERNODE_RPC,
+                    subsystem = LOG_SUBSYSTEM_NAMESPACE_SCANNER,
+                    operation = INTERNODE_OPERATION_NS_SCANNER,
+                    result = "failed",
+                    status_code = StatusCode::UPGRADE_REQUIRED.as_u16(),
+                    rpc_path = NS_SCANNER_PATH,
+                    method = %Method::GET,
+                    reason = "capability_authentication_unavailable",
+                    error = %err,
+                    "internode rpc request failed"
+                );
+                return response_with_status(StatusCode::UPGRADE_REQUIRED, "namespace scanner RPC authentication is unavailable");
+            }
+        };
     let body = match rmp_serde::to_vec_named(&NsScannerCapabilityResponse {
         version: NS_SCANNER_PROTOCOL_VERSION,
         server_epoch,
         proof,
+        supports_tier_registry_generation: include_tier_registry_generation.then_some(true),
     }) {
         Ok(body) => body,
         Err(err) => {
@@ -1676,12 +1685,13 @@ mod tests {
         LOG_SUBSYSTEM_NAMESPACE_SCANNER, LOG_SUBSYSTEM_ROUTING, NS_SCANNER_BODY_SHA256_QUERY,
         NS_SCANNER_CAPABILITY_CHALLENGE_QUERY, NS_SCANNER_CYCLE_QUERY, NS_SCANNER_LEADER_EPOCH_QUERY, NS_SCANNER_PATH,
         NS_SCANNER_REQUEST_ID_QUERY, NS_SCANNER_SERVER_EPOCH_QUERY, NS_SCANNER_SESSION_ID_QUERY,
-        NS_SCANNER_SESSION_SEQUENCE_QUERY, NsScannerQuery, PUT_FILE_AUTH_STREAM_PATH, PUT_FILE_CAPABILITY_PATH,
-        PUT_FILE_STREAM_PATH, PutFileQuery, READ_FILE_STREAM_PATH, WALK_DIR_BODY_SHA256_QUERY, WALK_DIR_PATH, WalkDirQuery,
-        append_walk_dir_completion, internode_http_operation, internode_rpc_subsystem, is_internode_rpc_path,
-        ns_scanner_response_body, ns_scanner_server_epoch_matches, put_body_size_mismatch, put_file_auth_nonce,
-        put_file_capability_response, put_file_server_epoch_matches, put_file_stage_error_message, put_file_target_lock,
-        read_file_body_stream, read_file_stream_buffer_size, remote_scanner_claim_rejection, response_with_disk_error,
+        NS_SCANNER_SESSION_SEQUENCE_QUERY, NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY, NsScannerCapabilityResponse,
+        NsScannerQuery, PUT_FILE_AUTH_STREAM_PATH, PUT_FILE_CAPABILITY_PATH, PUT_FILE_STREAM_PATH, PutFileQuery,
+        READ_FILE_STREAM_PATH, WALK_DIR_BODY_SHA256_QUERY, WALK_DIR_PATH, WalkDirQuery, append_walk_dir_completion,
+        internode_http_operation, internode_rpc_subsystem, is_internode_rpc_path, ns_scanner_response_body,
+        ns_scanner_server_epoch_matches, put_body_size_mismatch, put_file_auth_nonce, put_file_capability_response,
+        put_file_server_epoch_matches, put_file_stage_error_message, put_file_target_lock, read_file_body_stream,
+        read_file_stream_buffer_size, remote_scanner_claim_rejection, response_with_disk_error,
         supports_walk_dir_stream_completion, validate_walk_dir_completion_request, verify_internode_rpc_signature,
         verify_ns_scanner_body_digest, verify_walk_dir_body_digest, walk_dir_response_body, write_authenticated_put_file,
         write_body_chunks_to_writer, write_put_file_body_chunks_to_writer,
@@ -2114,6 +2124,51 @@ mod tests {
         );
         assert!(serde_urlencoded::from_str::<NsScannerQuery>(&query).is_err());
         assert!(serde_urlencoded::from_str::<super::NsScannerCapabilityQuery>("ns_scanner_protocol=1&unexpected=true").is_err());
+        let marked =
+            format!("ns_scanner_protocol=3&ns_scanner_challenge={request_id}&{NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY}=true");
+        assert_eq!(
+            serde_urlencoded::from_str::<super::NsScannerCapabilityQuery>(&marked)
+                .expect("generation marker should be accepted")
+                .ns_scanner_tier_registry_generation,
+            Some(true)
+        );
+    }
+
+    #[test]
+    fn namespace_scanner_capability_response_support_is_optional_for_old_peers() {
+        #[derive(serde::Deserialize)]
+        #[serde(deny_unknown_fields)]
+        struct LegacyCapabilityResponse {
+            version: u16,
+            server_epoch: uuid::Uuid,
+            proof: Vec<u8>,
+        }
+
+        let old = rmp_serde::to_vec_named(&NsScannerCapabilityResponse {
+            version: super::NS_SCANNER_PROTOCOL_VERSION,
+            server_epoch: uuid::Uuid::new_v4(),
+            proof: vec![1, 2, 3],
+            supports_tier_registry_generation: None,
+        })
+        .expect("old response shape should encode");
+        let decoded: NsScannerCapabilityResponse = rmp_serde::from_slice(&old).expect("old response should decode");
+        assert_eq!(decoded.supports_tier_registry_generation, None);
+        let legacy_decoded: LegacyCapabilityResponse =
+            rmp_serde::from_slice(&old).expect("legacy reader should decode old shape");
+        assert_eq!(legacy_decoded.version, super::NS_SCANNER_PROTOCOL_VERSION);
+        assert!(!legacy_decoded.server_epoch.is_nil());
+        assert_eq!(legacy_decoded.proof, vec![1, 2, 3]);
+
+        let current = rmp_serde::to_vec_named(&NsScannerCapabilityResponse {
+            version: super::NS_SCANNER_PROTOCOL_VERSION,
+            server_epoch: uuid::Uuid::new_v4(),
+            proof: vec![1, 2, 3],
+            supports_tier_registry_generation: Some(true),
+        })
+        .expect("new response shape should encode");
+        let decoded: NsScannerCapabilityResponse = rmp_serde::from_slice(&current).expect("new response should decode");
+        assert_eq!(decoded.supports_tier_registry_generation, Some(true));
+        assert!(rmp_serde::from_slice::<LegacyCapabilityResponse>(&current).is_err());
     }
 
     #[test]

rustfs/src/storage/storage_api.rs

@@ -224,8 +224,8 @@ pub(crate) mod rpc_consumer {
         pub(crate) use super::super::storage_contracts::{
             NS_SCANNER_BODY_SHA256_QUERY, NS_SCANNER_CAPABILITY_CHALLENGE_QUERY, NS_SCANNER_CYCLE_QUERY,
             NS_SCANNER_LEADER_EPOCH_QUERY, NS_SCANNER_REQUEST_ID_QUERY, NS_SCANNER_SERVER_EPOCH_QUERY,
-            NS_SCANNER_SESSION_ID_QUERY, NS_SCANNER_SESSION_SEQUENCE_QUERY, PUT_FILE_CAPABILITY_CHALLENGE_QUERY,
-            PUT_FILE_CAPABILITY_QUERY, WALK_DIR_BODY_SHA256_QUERY,
+            NS_SCANNER_SESSION_ID_QUERY, NS_SCANNER_SESSION_SEQUENCE_QUERY, NS_SCANNER_TIER_REGISTRY_GENERATION_QUERY,
+            PUT_FILE_CAPABILITY_CHALLENGE_QUERY, PUT_FILE_CAPABILITY_QUERY, WALK_DIR_BODY_SHA256_QUERY,
         };
         pub(crate) use super::super::storage_contracts::{
             NS_SCANNER_PROTOCOL_VERSION, NsScannerCapabilityResponse, PUT_FILE_AUTH_TRAILER_LEN, PUT_FILE_AUTH_V1,
@@ -233,8 +233,8 @@ pub(crate) mod rpc_consumer {
         };
         pub(crate) use super::super::{
             DeleteOptions, DiskStore, StorageDiskRpcExt, WalkDirOptions, check_and_record_signed_rpc_nonce,
-            find_local_disk_by_ref, sign_ns_scanner_capability, sign_put_file_capability, verify_put_file_auth_trailer,
-            verify_rpc_signature,
+            find_local_disk_by_ref, sign_ns_scanner_capability_with_tier_registry_generation, sign_put_file_capability,
+            verify_put_file_auth_trailer, verify_rpc_signature,
         };
     }
 
@@ -520,9 +520,10 @@ pub(crate) mod ecstore_rpc {
     pub(crate) use rustfs_ecstore::api::rpc::{
         KMS_SIGNAL_SUBSYSTEM, LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient,
         PeerS3Client, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC, TONIC_RPC_PREFIX,
-        check_and_record_signed_rpc_nonce, normalize_tonic_rpc_audience, sign_ns_scanner_capability, sign_put_file_capability,
-        sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers,
-        tonic_rpc_auth_failure_reason, verify_put_file_auth_trailer, verify_rpc_signature, verify_tonic_canonical_body_digest,
+        check_and_record_signed_rpc_nonce, normalize_tonic_rpc_audience,
+        sign_ns_scanner_capability_with_tier_registry_generation, sign_put_file_capability, sign_tonic_rpc_response_proof,
+        tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers, tonic_rpc_auth_failure_reason,
+        verify_put_file_auth_trailer, verify_rpc_signature, verify_tonic_canonical_body_digest,
         verify_tonic_mutation_body_digest, verify_tonic_rpc_signature_with_bootstrap,
     };
     #[cfg(test)]
@@ -1731,8 +1732,16 @@ pub(crate) fn verify_put_file_auth_trailer(
     ecstore_rpc::verify_put_file_auth_trailer(url, method, nonce, trailer)
 }
 
-pub(crate) fn sign_ns_scanner_capability(challenge: uuid::Uuid, server_epoch: uuid::Uuid) -> std::io::Result<Vec<u8>> {
-    ecstore_rpc::sign_ns_scanner_capability(challenge, server_epoch)
+pub(crate) fn sign_ns_scanner_capability_with_tier_registry_generation(
+    challenge: uuid::Uuid,
+    server_epoch: uuid::Uuid,
+    supports_tier_registry_generation: bool,
+) -> std::io::Result<Vec<u8>> {
+    ecstore_rpc::sign_ns_scanner_capability_with_tier_registry_generation(
+        challenge,
+        server_epoch,
+        supports_tier_registry_generation,
+    )
 }
 
 pub(crate) fn sign_put_file_capability(

scripts/test/vault_ha_kms_live.sh

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