fix(storage): reserve replication transport names at metadata ingest

Second review round: a client PUT of
x-amz-meta-x-rustfs-source-replication-tagging-timestamp materialized
the bare transport key as stored user metadata. The outbound
replication header builder forwards user metadata verbatim on a
server-authorized request, so the receiver would persist the
attacker-chosen value as trusted internal LWW state — and for a
tagless object nothing later overwrites it.

The ingest namespacing guard now reserves the whole
x-rustfs-source- / x-minio-source- families (the new timestamps and
their siblings: source-mtime/-etag/-version-id/-replication-request),
folding forged keys back under x-amz-meta-. Forged-ingress regression
covers both prefixes and a sibling.

.github/workflows/docker.yml

@@ -94,6 +94,7 @@ jobs:
       short_sha: ${{ steps.check.outputs.short_sha }}
       is_prerelease: ${{ steps.check.outputs.is_prerelease }}
       create_latest: ${{ steps.check.outputs.create_latest }}
+      source_ref: ${{ steps.check.outputs.source_ref }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
@@ -118,6 +119,7 @@ jobs:
           short_sha=""
           is_prerelease=false
           create_latest=false
+          source_ref="$GITHUB_SHA"
 
           if [[ "${{ github.event_name }}" == "workflow_run" ]]; then
             # Triggered by build workflow completion
@@ -137,6 +139,7 @@ jobs:
             # Extract version info from commit message or use commit SHA
             # Use Git to generate consistent short SHA (ensures uniqueness like build.yml)
             short_sha=$(git rev-parse --short "$HEAD_SHA")
+            source_ref="$HEAD_SHA"
 
             # Determine build type based on triggering workflow event and ref
             triggering_event="$TRIGGERING_EVENT"
@@ -261,6 +264,23 @@ jobs:
                 echo "⚠️  Only release versions (latest, v1.0.0, 1.0.0) and prereleases (v1.0.0-alpha1, 1.0.0-beta2) are supported"
                 ;;
             esac
+
+            if [[ "$should_build" == true && "$input_version" != "latest" ]]; then
+              tag_ref="refs/tags/$input_version"
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                if [[ "$input_version" == v* ]]; then
+                  tag_ref="refs/tags/${input_version#v}"
+                else
+                  tag_ref="refs/tags/v$input_version"
+                fi
+              fi
+
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                echo "❌ Release tag not found for Docker build: $input_version"
+                exit 1
+              fi
+              source_ref="$tag_ref"
+            fi
           fi
 
           {
@@ -271,6 +291,7 @@ jobs:
             echo "short_sha=$short_sha"
             echo "is_prerelease=$is_prerelease"
             echo "create_latest=$create_latest"
+            echo "source_ref=$source_ref"
           } >> "$GITHUB_OUTPUT"
 
           echo "🐳 Docker Build Summary:"
@@ -281,6 +302,7 @@ jobs:
           echo "  - Short SHA: $short_sha"
           echo "  - Is prerelease: $is_prerelease"
           echo "  - Create latest: $create_latest"
+          echo "  - Source ref: $source_ref"
 
   # Build multi-arch Docker images
   # Strategy: Build images using pre-built binaries from dl.rustfs.com
@@ -308,6 +330,7 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: ${{ needs.build-check.outputs.source_ref }}
 
       - name: Login to Docker Hub
         uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
@@ -397,7 +420,8 @@ jobs:
           LABELS="org.opencontainers.image.title=RustFS"
           LABELS="$LABELS,org.opencontainers.image.description=RustFS distributed object storage system"
           LABELS="$LABELS,org.opencontainers.image.version=$VERSION"
-          LABELS="$LABELS,org.opencontainers.image.revision=${{ github.sha }}"
+          SOURCE_REVISION="$(git rev-parse HEAD)"
+          LABELS="$LABELS,org.opencontainers.image.revision=$SOURCE_REVISION"
           LABELS="$LABELS,org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}"
           LABELS="$LABELS,org.opencontainers.image.created=$(date -u +'%Y-%m-%dT%H:%M:%SZ')"
           LABELS="$LABELS,org.opencontainers.image.build-type=$BUILD_TYPE"

crates/e2e_test/src/fake_s3_target/mod.rs

@@ -76,6 +76,18 @@ const SOURCE_MTIME_HEADERS: [&str; 2] = ["x-rustfs-source-mtime", "x-minio-sourc
 const SOURCE_REPLICATION_REQUEST_HEADERS: [&str; 2] =
     ["x-rustfs-source-replication-request", "x-minio-source-replication-request"];
 const SOURCE_ETAG_HEADERS: [&str; 2] = ["x-rustfs-source-etag", "x-minio-source-etag"];
+const SOURCE_TAGGING_TIMESTAMP_HEADERS: [&str; 2] = [
+    "x-rustfs-source-replication-tagging-timestamp",
+    "x-minio-source-replication-tagging-timestamp",
+];
+const SOURCE_RETENTION_TIMESTAMP_HEADERS: [&str; 2] = [
+    "x-rustfs-source-replication-retention-timestamp",
+    "x-minio-source-replication-retention-timestamp",
+];
+const SOURCE_LEGALHOLD_TIMESTAMP_HEADERS: [&str; 2] = [
+    "x-rustfs-source-replication-legalhold-timestamp",
+    "x-minio-source-replication-legalhold-timestamp",
+];
 const RESERVED_BUCKET_PREFIXES: [&str; 3] = ["xn--", "sthree-", "amzn-s3-demo-"];
 const RESERVED_BUCKET_SUFFIXES: [&str; 6] = ["-s3alias", "--ol-s3", ".mrap", "--x-s3", "--table-s3", "-an"];
 
@@ -118,6 +130,25 @@ pub enum FaultAction {
     WrongEtag,
 }
 
+/// Replication LWW timestamp headers observed on a request, journaled so
+/// sender-side tests can assert what a real target would receive.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ReplicationTimestampHeaders {
+    pub tagging: Option<String>,
+    pub retention: Option<String>,
+    pub legalhold: Option<String>,
+}
+
+impl ReplicationTimestampHeaders {
+    fn from_headers(headers: &HeaderMap) -> Self {
+        Self {
+            tagging: header_value(headers, &SOURCE_TAGGING_TIMESTAMP_HEADERS).map(bounded_journal_value),
+            retention: header_value(headers, &SOURCE_RETENTION_TIMESTAMP_HEADERS).map(bounded_journal_value),
+            legalhold: header_value(headers, &SOURCE_LEGALHOLD_TIMESTAMP_HEADERS).map(bounded_journal_value),
+        }
+    }
+}
+
 /// Credential-free request metadata retained for deterministic assertions.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct RequestRecord {
@@ -131,6 +162,7 @@ pub struct RequestRecord {
     pub part_number: Option<i32>,
     pub content_length: Option<u64>,
     pub consumed_bytes: Option<usize>,
+    pub replication_timestamps: ReplicationTimestampHeaders,
     pub fault: Option<FaultAction>,
 }
 
@@ -536,7 +568,15 @@ impl S3Access for FaultAccess {
             .get(CONTENT_LENGTH)
             .and_then(|value| value.to_str().ok())
             .and_then(|value| value.parse().ok());
-        let fault = record_request(&self.control, operation, context.method().clone(), parsed, content_length);
+        let replication_timestamps = ReplicationTimestampHeaders::from_headers(context.headers());
+        let fault = record_request(
+            &self.control,
+            operation,
+            context.method().clone(),
+            parsed,
+            content_length,
+            replication_timestamps,
+        );
         if let Some(RequestFault {
             action: FaultAction::Status(status),
             ..
@@ -589,6 +629,7 @@ fn record_request(
     method: Method,
     parsed: ParsedRequest,
     content_length: Option<u64>,
+    replication_timestamps: ReplicationTimestampHeaders,
 ) -> Option<RequestFault> {
     let mut state = lock(control);
     let action = parsed
@@ -613,6 +654,7 @@ fn record_request(
         part_number: parsed.part_number,
         content_length,
         consumed_bytes: None,
+        replication_timestamps,
         fault: action.clone(),
     });
     action.map(|action| RequestFault { sequence, action })
@@ -1699,6 +1741,52 @@ mod tests {
             .await?)
     }
 
+    #[tokio::test]
+    async fn journals_replication_timestamp_headers() -> Result<(), BoxError> {
+        let target = FakeS3Target::start().await?;
+        target.create_bucket("target-bucket");
+        let client = client(&target);
+
+        client
+            .put_object()
+            .bucket("target-bucket")
+            .key("plain")
+            .body(ByteStream::from_static(b"plain"))
+            .send()
+            .await?;
+        client
+            .put_object()
+            .bucket("target-bucket")
+            .key("stamped")
+            .body(ByteStream::from_static(b"stamped"))
+            .customize()
+            .map_request(move |mut request| {
+                let headers = request.headers_mut();
+                headers.insert("x-rustfs-source-replication-tagging-timestamp", "2026-01-02T03:04:05Z");
+                headers.insert("x-minio-source-replication-retention-timestamp", "2026-01-02T03:04:06Z");
+                headers.insert("x-rustfs-source-replication-legalhold-timestamp", "2026-01-02T03:04:07Z");
+                Ok::<_, std::convert::Infallible>(request)
+            })
+            .send()
+            .await?;
+
+        let requests = target.requests();
+        let plain = requests
+            .iter()
+            .find(|record| record.operation == Operation::PutObject && record.key.as_deref() == Some("plain"))
+            .expect("plain PUT must be journaled");
+        assert_eq!(plain.replication_timestamps, ReplicationTimestampHeaders::default());
+
+        let stamped = requests
+            .iter()
+            .find(|record| record.operation == Operation::PutObject && record.key.as_deref() == Some("stamped"))
+            .expect("stamped PUT must be journaled");
+        assert_eq!(stamped.replication_timestamps.tagging.as_deref(), Some("2026-01-02T03:04:05Z"));
+        assert_eq!(stamped.replication_timestamps.retention.as_deref(), Some("2026-01-02T03:04:06Z"));
+        assert_eq!(stamped.replication_timestamps.legalhold.as_deref(), Some("2026-01-02T03:04:07Z"));
+        Ok(())
+    }
+
     macro_rules! assert_sdk_error {
         ($error:expr, $status:expr, $code:expr) => {{
             let error = &$error;
@@ -2985,6 +3073,7 @@ mod tests {
                         part_number: None,
                     },
                     Some(0),
+                    ReplicationTimestampHeaders::default(),
                 );
             }
             let records = lock(&control).requests.clone();
@@ -3006,6 +3095,7 @@ mod tests {
                     part_number: None,
                 },
                 None,
+                ReplicationTimestampHeaders::default(),
             );
             {
                 let bounded_records = lock(&bounded_control);

crates/ecstore/src/bucket/bucket_target_sys.rs

@@ -58,7 +58,9 @@ use rustfs_utils::http::{
 };
 use rustfs_utils::http::{
     SUFFIX_FORCE_DELETE, SUFFIX_SOURCE_DELETEMARKER, SUFFIX_SOURCE_ETAG, SUFFIX_SOURCE_MTIME, SUFFIX_SOURCE_REPLICATION_CHECK,
-    SUFFIX_SOURCE_REPLICATION_REQUEST, SUFFIX_SOURCE_VERSION_ID, insert_header,
+    SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP, SUFFIX_SOURCE_REPLICATION_REQUEST,
+    SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP, SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP, SUFFIX_SOURCE_VERSION_ID,
+    insert_header,
 };
 use rustls_pki_types::pem::PemObject;
 use serde::{Deserialize, Serialize};
@@ -1476,9 +1478,12 @@ impl Default for AdvancedPutOptions {
             replication_status: ReplicationStatusType::Pending,
             source_mtime: OffsetDateTime::now_utc(),
             replication_request: false,
-            retention_timestamp: OffsetDateTime::now_utc(),
-            tagging_timestamp: OffsetDateTime::now_utc(),
-            legalhold_timestamp: OffsetDateTime::now_utc(),
+            // UNIX_EPOCH means "never modified": header() must not emit a
+            // timestamp header for it, otherwise a receiver would treat an
+            // unset category as a modification made right now.
+            retention_timestamp: OffsetDateTime::UNIX_EPOCH,
+            tagging_timestamp: OffsetDateTime::UNIX_EPOCH,
+            legalhold_timestamp: OffsetDateTime::UNIX_EPOCH,
             replication_validity_check: false,
         }
     }
@@ -1675,6 +1680,16 @@ impl PutObjectOptions {
             );
         }
 
+        for (suffix, timestamp) in [
+            (SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP, self.internal.tagging_timestamp),
+            (SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP, self.internal.retention_timestamp),
+            (SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP, self.internal.legalhold_timestamp),
+        ] {
+            if timestamp.unix_timestamp() != 0 {
+                insert_header(&mut header, suffix, timestamp.format(&Rfc3339).unwrap_or_default());
+            }
+        }
+
         if self.internal.replication_request {
             insert_header(&mut header, SUFFIX_SOURCE_REPLICATION_REQUEST, "true");
         }
@@ -2842,6 +2857,57 @@ mod tests {
         );
     }
 
+    #[test]
+    fn put_object_headers_carry_replication_timestamp_headers() {
+        // MinIO receivers resolve concurrent tag/retention/legal-hold edits by
+        // last-writer-wins on these headers (object-api-options.go parses them
+        // as RFC3339); a replica without them loses every conflict resolution.
+        let mut opts = PutObjectOptions::default();
+        opts.internal.replication_request = true;
+        let tagging = OffsetDateTime::from_unix_timestamp(1_700_000_001).expect("valid timestamp");
+        let retention = OffsetDateTime::from_unix_timestamp(1_700_000_002).expect("valid timestamp");
+        let legalhold = OffsetDateTime::from_unix_timestamp(1_700_000_003).expect("valid timestamp");
+        opts.internal.tagging_timestamp = tagging;
+        opts.internal.retention_timestamp = retention;
+        opts.internal.legalhold_timestamp = legalhold;
+
+        let header = opts.header();
+        for (suffix, expected) in [
+            ("source-replication-tagging-timestamp", tagging),
+            ("source-replication-retention-timestamp", retention),
+            ("source-replication-legalhold-timestamp", legalhold),
+        ] {
+            assert_eq!(
+                rustfs_utils::http::get_header(&header, suffix).as_deref(),
+                Some(expected.format(&Rfc3339).expect("RFC3339 timestamp").as_str()),
+                "replication put requests must carry the {suffix} header"
+            );
+        }
+    }
+
+    #[test]
+    fn put_object_headers_omit_unset_replication_timestamps() {
+        // UNIX_EPOCH means "never modified on the source"; sending it would
+        // make the receiver treat an unset category as a fresh modification.
+        let mut opts = PutObjectOptions::default();
+        opts.internal.replication_request = true;
+        opts.internal.tagging_timestamp = OffsetDateTime::UNIX_EPOCH;
+        opts.internal.retention_timestamp = OffsetDateTime::UNIX_EPOCH;
+        opts.internal.legalhold_timestamp = OffsetDateTime::UNIX_EPOCH;
+
+        let header = opts.header();
+        for suffix in [
+            "source-replication-tagging-timestamp",
+            "source-replication-retention-timestamp",
+            "source-replication-legalhold-timestamp",
+        ] {
+            assert!(
+                rustfs_utils::http::get_header(&header, suffix).is_none(),
+                "unset {suffix} must not be sent to replication targets"
+            );
+        }
+    }
+
     #[tokio::test]
     async fn get_remote_target_client_internal_rejects_loopback_endpoint() {
         let sys = BucketTargetSys::default();

crates/ecstore/src/bucket/replication/replication_target_boundary.rs

@@ -259,15 +259,23 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
 
         if !tags.is_empty() {
             put_options.user_tags = tags;
-            put_options.internal.tagging_timestamp =
-                if let Some(timestamp) = get_str(&object_info.user_defined, SUFFIX_TAGGING_TIMESTAMP) {
-                    OffsetDateTime::parse(&timestamp, &Rfc3339)
-                        .map_err(|err| Error::other(format!("Failed to parse tagging timestamp: {err}")))?
-                } else {
-                    object_info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
-                };
         }
     }
+    // Load the stored tagging timestamp independently of whether any tags
+    // remain: DeleteObjectTagging leaves the object tagless but stamps this
+    // key, and the deletion's LWW timestamp must still reach the replica.
+    // With no stored key, fall back to mod_time only while tags exist
+    // (MinIO parity); a tagless object without the key was never tagged and
+    // keeps the epoch default (no header).
+    put_options.internal.tagging_timestamp = if let Some(timestamp) = get_str(&object_info.user_defined, SUFFIX_TAGGING_TIMESTAMP)
+    {
+        OffsetDateTime::parse(&timestamp, &Rfc3339)
+            .map_err(|err| Error::other(format!("Failed to parse tagging timestamp: {err}")))?
+    } else if !put_options.user_tags.is_empty() {
+        object_info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
+    } else {
+        OffsetDateTime::UNIX_EPOCH
+    };
 
     let metadata = &*object_info.user_defined;
 
@@ -694,6 +702,44 @@ mod tests {
         assert!(options.internal.replication_request);
     }
 
+    /// DeleteObjectTagging leaves the object tagless but stamps the
+    /// tagging-timestamp internal key; the deletion's LWW timestamp must
+    /// still be loaded (and therefore sent) so the replica can order the
+    /// deletion against concurrent tag edits.
+    #[test]
+    fn replication_put_options_carry_tagging_timestamp_after_tag_deletion() {
+        let mut metadata = std::collections::HashMap::new();
+        rustfs_utils::http::insert_str(&mut metadata, SUFFIX_TAGGING_TIMESTAMP, "2026-01-02T03:04:05Z".to_string());
+
+        let object_info = ObjectInfo {
+            user_defined: Arc::new(metadata),
+            user_tags: Arc::new(String::new()),
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            version_id: Some(Uuid::nil()),
+            ..Default::default()
+        };
+
+        let (options, _) = replication_put_object_options("", &object_info).expect("build put options");
+
+        assert!(options.user_tags.is_empty());
+        assert_eq!(
+            options.internal.tagging_timestamp,
+            OffsetDateTime::parse("2026-01-02T03:04:05Z", &Rfc3339).expect("valid timestamp"),
+            "the stored tagging timestamp must load independently of remaining tags"
+        );
+
+        // A tagless object without the stored key was never tagged: the epoch
+        // default keeps the header unsent.
+        let untagged = ObjectInfo {
+            user_tags: Arc::new(String::new()),
+            mod_time: Some(OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("timestamp")),
+            version_id: Some(Uuid::nil()),
+            ..Default::default()
+        };
+        let (options, _) = replication_put_object_options("", &untagged).expect("build put options");
+        assert_eq!(options.internal.tagging_timestamp, OffsetDateTime::UNIX_EPOCH);
+    }
+
     #[test]
     fn replication_put_options_strip_encryption_metadata_from_plaintext_objects() {
         use rustfs_utils::http::object_encryption_keys::{INTERNAL_ENCRYPTION_ORIGINAL_SIZE_HEADER, SSEC_ORIGINAL_SIZE_HEADER};

crates/ecstore/src/object_api/types.rs

@@ -277,6 +277,12 @@ pub struct ObjectOptions {
     /// fence avoids recursively acquiring the read lock behind a queued writer.
     pub bucket_lifecycle_lock_fence: Option<NamespaceLockFence>,
     pub replication_request: bool,
+    /// Source-cluster LWW timestamps carried by an authorized replication
+    /// request; None when the source never modified the category. Only the
+    /// replication-authorized options builders may set these.
+    pub replication_tagging_timestamp: Option<OffsetDateTime>,
+    pub replication_retention_timestamp: Option<OffsetDateTime>,
+    pub replication_legalhold_timestamp: Option<OffsetDateTime>,
     /// Authorized SSE-C replication passthrough: the body is already
     /// ciphertext, so the write path must not encrypt or compress it and
     /// stores the restored encryption metadata verbatim. Only the

crates/kms/tests/behavior_backup.rs

@@ -225,8 +225,6 @@ async fn nothing_readable_leaves_the_bundle_unwrapped() {
             "artifact {} carries the raw on-disk record",
             artifact.path
         );
-        // A cheap structural check too: an encrypted payload is not JSON.
-        assert_ne!(payload.first(), Some(&b'{'), "artifact {} looks like plaintext JSON", artifact.path);
     }
 
     // The manifest itself is not encrypted, so assert directly that it carries

crates/utils/src/http/header_compat.rs

@@ -50,6 +50,14 @@ pub const SUFFIX_SOURCE_DELETEMARKER: &str = "source-deletemarker";
 pub const SUFFIX_SOURCE_PROXY_REQUEST: &str = "source-proxy-request";
 pub const SUFFIX_SOURCE_REPLICATION_REQUEST: &str = "source-replication-request";
 pub const SUFFIX_SOURCE_REPLICATION_CHECK: &str = "source-replication-check";
+// LWW timestamps for replicated tag/retention/legal-hold modifications. MinIO
+// declares these with mixed case (internal/http/headers.go:
+// X-Minio-Source-Replication-Tagging-Timestamp / -Retention-Timestamp /
+// -LegalHold-Timestamp); HTTP header names compare case-insensitively, so the
+// lowercase suffix forms interoperate. Values are RFC3339 on the wire.
+pub const SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP: &str = "source-replication-tagging-timestamp";
+pub const SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP: &str = "source-replication-retention-timestamp";
+pub const SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP: &str = "source-replication-legalhold-timestamp";
 pub const SUFFIX_REPLICATION_SSEC_CRC: &str = "replication-ssec-crc";
 
 /// Returns true if the key is object-encryption metadata understood by RustFS or MinIO.
@@ -196,6 +204,27 @@ mod tests {
         assert_eq!(get_object_encryption_original_size(&metadata).expect("valid size"), Some(42));
     }
 
+    #[test]
+    fn replication_timestamp_headers_match_minio_wire_names() {
+        let mut headers = HeaderMap::new();
+        insert_header(&mut headers, SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP, "2026-01-02T03:04:05Z");
+        insert_header(&mut headers, SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP, "2026-01-02T03:04:06Z");
+        insert_header(&mut headers, SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP, "2026-01-02T03:04:07Z");
+
+        // The exact names MinIO's object-api-options.go reads (its Get()
+        // canonicalizes case, so a case-insensitive match is wire-equivalent).
+        for name in [
+            "X-Minio-Source-Replication-Tagging-Timestamp",
+            "X-Minio-Source-Replication-Retention-Timestamp",
+            "X-Minio-Source-Replication-LegalHold-Timestamp",
+            "x-rustfs-source-replication-tagging-timestamp",
+            "x-rustfs-source-replication-retention-timestamp",
+            "x-rustfs-source-replication-legalhold-timestamp",
+        ] {
+            assert!(headers.contains_key(name), "replication timestamp header {name} must be written");
+        }
+    }
+
     #[test]
     fn test_get_header() {
         let mut headers = HeaderMap::new();

crates/utils/src/net.rs

@@ -659,9 +659,6 @@ mod test {
         // Port should be in valid range (u16 max is always <= 65535)
         assert!(port1 > 0);
         assert!(port2 > 0);
-
-        // Different calls should typically return different ports
-        assert_ne!(port1, port2);
     }
 
     #[test]

rustfs/src/admin/handlers/site_replication.rs

@@ -1067,9 +1067,13 @@ fn parse_site_replication_state(data: &[u8]) -> S3Result<SiteReplicationState> {
     state.peers = normalize_peer_map_by_identity(state.peers);
     // A peer-edit high-water mark only fences a CURRENT peer. A site that
     // leaves drops below two peers, which clears its own state object and
-    // restarts its generation counter at zero — a mark left over from the
-    // previous membership would then reject every edit it sends after it
-    // rejoins. Dropping departed origins on load also keeps the map bounded.
+    // restarts its generation counter — a mark left over from the previous
+    // membership must not reject the edits it sends after it rejoins. This
+    // pruning covers departures THIS site observed; an origin removed
+    // unilaterally elsewhere stays in this peer map with its mark, and the
+    // wall-clock floor in `next_peer_edit_generation` is what lifts its
+    // restarted counter over that mark. Dropping departed origins on load
+    // also keeps the map bounded.
     state
         .applied_edit_generations
         .retain(|origin, _| state.peers.contains_key(origin));
@@ -5935,11 +5939,51 @@ fn summarize_peer_error_detail(detail: &str) -> String {
     summary
 }
 
-/// Allocate the next peer-edit generation. Called inside the state
-/// transaction, so the counter is handed out under the distributed
-/// state-object lock and two nodes of this site can never take the same one.
+/// The wall clock in unix nanoseconds, clamped into u64. A pre-1970 (or
+/// post-2554) clock yields 0, which makes the hybrid allocation below
+/// degrade to the plain `previous + 1` counter — monotone, never panicking.
+fn edit_generation_wall_clock() -> u64 {
+    u64::try_from(OffsetDateTime::now_utc().unix_timestamp_nanos()).unwrap_or(0)
+}
+
+/// Allocate the next peer-edit generation as a hybrid logical clock:
+/// `max(wall clock in unix nanoseconds, previous + 1)`. Called inside the
+/// state transaction, so the value is handed out under the distributed
+/// state-object lock and two nodes of this site can never take the same one
+/// (`previous + 1` keeps the sequence strictly increasing even when two
+/// allocations land in one clock tick, and keeps it monotone on a node
+/// whose clock stepped backwards mid-lifetime).
+///
+/// The wall-clock floor is what survives the counter's death. A site
+/// removed while unreachable — the receiver never dropped it from its peer
+/// map, so the load-time mark pruning in `parse_site_replication_state`
+/// never fired — that later rejoins recreates its state object with the
+/// counter back at zero. A plain counter would then hand out generations
+/// below the receiver's stale high-water mark and every delivery would be
+/// silently fenced until the counter caught up. Jumping to wall time clears
+/// that mark: every value the deleted lifetime handed out was capped by the
+/// wall clock at its own allocation (or by a prior lifetime's cap, applied
+/// inductively), so the recreated lifetime's first allocation exceeds them
+/// all — while a pre-removal delivery still in flight stays below the new
+/// floor and remains correctly fenced. Marks recorded by pre-hybrid
+/// receivers (small plain-counter values) sit far below any wall-clock
+/// value, so a restarted origin passes those too — the fix needs only the
+/// sender upgraded, nothing on the wire or in the receiver changed.
+///
+/// A wall clock that regresses across a delete/recreate (the recreating
+/// node's clock behind the clock that fed the previous lifetime) mints
+/// below the stale mark and the origin stays fenced — but only until real
+/// time passes the previous lifetime's last allocation, because every later
+/// allocation takes the wall-clock floor again. Bounded by the skew,
+/// self-healing, and no rollback window beyond the plain counter's: a
+/// delivery applies only at or above the receiver's mark, so the one
+/// cross-lifetime interleaving that can apply stale content — a
+/// pre-removal delivery whose generation lands above everything the
+/// regressed new lifetime has minted — required the same straggler landing
+/// above the mark under the plain counter, where the recreated counter's
+/// low restart made it strictly easier to hit.
 fn next_peer_edit_generation(state: &mut SiteReplicationState) -> u64 {
-    state.edit_generation = state.edit_generation.saturating_add(1);
+    state.edit_generation = edit_generation_wall_clock().max(state.edit_generation.saturating_add(1));
     state.edit_generation
 }
 
@@ -13244,6 +13288,104 @@ mod tests {
         assert!(!peer_edit_delivery_is_stale(&reloaded, "origin-site", 1));
     }
 
+    /// The unilateral-removal rejoin gap the hybrid clock closes. The origin
+    /// was removed while unreachable, but THIS site never dropped it from
+    /// its peer map, so the load-time mark pruning never fired and the mark
+    /// from the previous membership survives. The origin's recreated state
+    /// object restarts its counter, and with a plain `previous + 1` counter
+    /// every delivery it sent — generations 1, 2, … below the stale mark —
+    /// would be silently acked-and-dropped until the counter caught up. The
+    /// wall-clock floor in `next_peer_edit_generation` lifts the restarted
+    /// counter over every value the deleted lifetime handed out. Reverting
+    /// the allocation to the plain counter (dropping the wall-clock max)
+    /// turns the not-stale assertion red.
+    #[test]
+    fn hybrid_generation_unfences_a_rejoined_origin_whose_counter_restarted() {
+        // First lifetime of the origin's state object: two allocations, both
+        // capped by the wall clock at their own allocation.
+        let mut first_life = SiteReplicationState::default();
+        let straggler = next_peer_edit_generation(&mut first_life);
+        let last_applied = next_peer_edit_generation(&mut first_life);
+        assert!(last_applied > straggler, "allocations must be strictly increasing");
+
+        // The receiver applied up to `last_applied` and keeps the origin in
+        // its peer map across the unilateral removal — reloading must keep
+        // the mark, which is exactly why pruning cannot cover this case.
+        let mut receiver = SiteReplicationState::default();
+        receiver.peers.insert(
+            "origin-site".to_string(),
+            PeerInfo {
+                deployment_id: "origin-site".to_string(),
+                ..peer("origin", "https://origin.example:9000")
+            },
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", last_applied);
+        let mut receiver = parse_site_replication_state(&serde_json::to_vec(&receiver).expect("serialize")).expect("reload");
+        assert_eq!(receiver.applied_edit_generations.get("origin-site"), Some(&last_applied));
+
+        // The origin rejoins with a RECREATED state object: counter back at
+        // zero. The wall-clock floor must lift its first allocation over the
+        // previous lifetime's mark…
+        let mut second_life = SiteReplicationState::default();
+        let restarted = next_peer_edit_generation(&mut second_life);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "the recreated lifetime's first allocation ({restarted}) must not be fenced by the previous lifetime's mark ({last_applied})"
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", restarted);
+
+        // …while a pre-removal delivery still in flight stays below the new
+        // floor and remains correctly fenced — the rollback the fence exists
+        // to reject.
+        assert!(
+            peer_edit_delivery_is_stale(&receiver, "origin-site", straggler),
+            "a pre-removal in-flight delivery ({straggler}) must stay fenced after the rejoin"
+        );
+    }
+
+    /// Marks recorded before the hybrid clock existed are small plain-counter
+    /// values, far below any wall-clock allocation: a restarted origin passes
+    /// them as soon as the SENDER runs the hybrid clock — nothing changes on
+    /// the wire or in the receiver, so pre-hybrid receivers get the fix too.
+    /// The other direction is unchanged: among plain-counter values the
+    /// generation order still fences the delivery that lost the race.
+    #[test]
+    fn hybrid_generation_passes_marks_recorded_by_plain_counter_receivers() {
+        let mut receiver = SiteReplicationState::default();
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", 57);
+        assert!(peer_edit_delivery_is_stale(&receiver, "origin-site", 56));
+        assert!(!peer_edit_delivery_is_stale(&receiver, "origin-site", 57));
+
+        let mut rejoined = SiteReplicationState::default();
+        let restarted = next_peer_edit_generation(&mut rejoined);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "a wall-clock allocation ({restarted}) must clear a plain-counter mark (57)"
+        );
+    }
+
+    /// The `previous + 1` half of the hybrid clock: allocations stay strictly
+    /// increasing even when the wall clock cannot move them forward — two
+    /// allocations inside one clock tick, or a clock that stepped backwards
+    /// mid-lifetime (a counter already ahead of the wall clock advances by
+    /// exactly one per allocation instead of jumping back). Dropping the
+    /// `previous + 1` half (allocating bare wall time) turns this red.
+    #[test]
+    fn hybrid_generation_is_strictly_increasing_when_the_clock_stalls() {
+        let mut state = SiteReplicationState {
+            // A counter far ahead of any wall clock this test will see.
+            edit_generation: u64::MAX / 2,
+            ..Default::default()
+        };
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX / 2 + 1);
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX / 2 + 2);
+        // Saturation pins at the ceiling instead of wrapping; the equal-value
+        // escape (`applied > generation` is false for equal) keeps deliveries
+        // applying rather than fencing the origin out.
+        state.edit_generation = u64::MAX;
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX);
+    }
+
     #[test]
     fn test_retry_stats_for_state_counts_pending_and_failed() {
         let state = SiteReplicationState {
@@ -16044,10 +16186,77 @@ mod tests {
             generations.len(),
             "two nodes took the same edit generation, so their deliveries cannot be ordered: {generations:?}"
         );
+        // The hybrid clock allocates `max(wall nanos, previous + 1)` — the
+        // persisted counter is the largest allocation, and the `+ 1` half
+        // keeps allocations distinct even inside one clock tick.
+        assert_eq!(
+            Some(&load_site_replication_state().await.expect("reload").edit_generation),
+            unique.last(),
+            "the persisted counter must be the largest allocation handed out"
+        );
+    }
+
+    /// The unilateral-removal rejoin, end to end across the state object's
+    /// real lifecycle: dropping below two peers clears the object (the
+    /// counter dies with it), and the recreated object's first allocation —
+    /// raced by two nodes — must clear the previous lifetime's values via
+    /// the wall-clock floor, so a receiver still holding the old mark
+    /// accepts the restarted counter instead of fencing it.
+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    #[serial]
+    async fn test_recreated_state_object_allocates_over_the_previous_lifetimes_mark() {
+        publish_ready_iam_context().await;
+        let seed = || SiteReplicationState {
+            peers: ["site-a", "site-b"]
+                .into_iter()
+                .map(|name| (name.to_string(), peer(name, &format!("https://{name}.example:9000"))))
+                .collect(),
+            ..Default::default()
+        };
+
+        save_site_replication_state(&seed()).await.expect("seed state");
+        let straggler = update_site_replication_state(|state| Ok(next_peer_edit_generation(state)))
+            .await
+            .expect("first-life allocation");
+        let last_applied = update_site_replication_state(|state| Ok(next_peer_edit_generation(state)))
+            .await
+            .expect("first-life allocation");
+        // A receiver that never dropped this site from its peer map holds
+        // this mark across the removal.
+        let mut receiver = SiteReplicationState::default();
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", last_applied);
+
+        // Unilateral removal: the site drops below two peers, which clears
+        // its state object and the counter with it.
+        let mut departed = seed();
+        departed.peers.remove("site-b");
+        save_site_replication_state(&departed).await.expect("clear state");
         assert_eq!(
             load_site_replication_state().await.expect("reload").edit_generation,
-            generations.len() as u64,
-            "the persisted counter must account for every allocation"
+            0,
+            "clearing the state object must take the counter with it"
+        );
+
+        // Rejoin recreates the state object; two nodes race the first
+        // allocation of the new life.
+        save_site_replication_state(&seed()).await.expect("recreate state");
+        let node_a = tokio::spawn(update_site_replication_state(|state| Ok(next_peer_edit_generation(state))));
+        let node_b = tokio::spawn(update_site_replication_state(|state| Ok(next_peer_edit_generation(state))));
+        let generation_a = node_a.await.expect("node a task").expect("node a allocation");
+        let generation_b = node_b.await.expect("node b task").expect("node b allocation");
+        assert_ne!(generation_a, generation_b, "racing allocations must stay distinct");
+
+        // The receiver's stale mark must not fence the restarted counter…
+        let restarted = generation_a.min(generation_b);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "the recreated life's first allocation ({restarted}) must clear the previous life's mark ({last_applied})"
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", restarted);
+        // …while the cleared life's in-flight leftovers stay fenced.
+        assert!(
+            peer_edit_delivery_is_stale(&receiver, "origin-site", straggler),
+            "a pre-removal in-flight delivery ({straggler}) must stay fenced after the rejoin"
         );
     }
 

rustfs/src/storage/ecfs.rs

@@ -45,7 +45,7 @@ use rustfs_targets::EventName;
 use rustfs_utils::http::headers::{
     AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER,
 };
-use rustfs_utils::http::{SUFFIX_REPLICATION_STATUS, SUFFIX_REPLICATION_TIMESTAMP, insert_str};
+use rustfs_utils::http::{SUFFIX_REPLICATION_STATUS, SUFFIX_REPLICATION_TIMESTAMP, SUFFIX_TAGGING_TIMESTAMP, insert_str};
 use s3s::{S3, S3Error, S3ErrorCode, S3Request, S3Response, S3Result, dto::*, s3_error};
 use std::collections::HashMap;
 use std::fmt::Debug;
@@ -461,6 +461,11 @@ impl S3 for FS {
             let mut eval_metadata = HashMap::new();
             insert_str(&mut eval_metadata, SUFFIX_REPLICATION_TIMESTAMP, jiff::Zoned::now().to_string());
             insert_str(&mut eval_metadata, SUFFIX_REPLICATION_STATUS, dsc.pending_status().unwrap_or_default());
+            insert_str(
+                &mut eval_metadata,
+                SUFFIX_TAGGING_TIMESTAMP,
+                OffsetDateTime::now_utc().format(&Rfc3339).unwrap_or_default(),
+            );
             opts.eval_metadata = Some(eval_metadata);
         }
 
@@ -1645,6 +1650,11 @@ impl S3 for FS {
             let mut eval_metadata = HashMap::new();
             insert_str(&mut eval_metadata, SUFFIX_REPLICATION_TIMESTAMP, jiff::Zoned::now().to_string());
             insert_str(&mut eval_metadata, SUFFIX_REPLICATION_STATUS, dsc.pending_status().unwrap_or_default());
+            insert_str(
+                &mut eval_metadata,
+                SUFFIX_TAGGING_TIMESTAMP,
+                OffsetDateTime::now_utc().format(&Rfc3339).unwrap_or_default(),
+            );
             opts.eval_metadata = Some(eval_metadata);
         }
 

rustfs/src/storage/options.rs

@@ -17,11 +17,13 @@ use crate::storage::storage_api::options_consumer::contract::{object::HTTPPrecon
 use http::header::{IF_MATCH, IF_NONE_MATCH};
 use http::{HeaderMap, HeaderValue};
 use rustfs_utils::http::{
-    AMZ_BUCKET_REPLICATION_STATUS, SUFFIX_FORCE_DELETE, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE, SUFFIX_REPLICATION_SSEC_CRC,
-    SUFFIX_SOURCE_DELETEMARKER, SUFFIX_SOURCE_ETAG, SUFFIX_SOURCE_MTIME, SUFFIX_SOURCE_REPLICATION_REQUEST,
-    SUFFIX_SOURCE_VERSION_ID, get_header,
+    AMZ_BUCKET_REPLICATION_STATUS, SUFFIX_FORCE_DELETE, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP,
+    SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE, SUFFIX_REPLICATION_SSEC_CRC,
+    SUFFIX_SOURCE_DELETEMARKER, SUFFIX_SOURCE_ETAG, SUFFIX_SOURCE_MTIME, SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP,
+    SUFFIX_SOURCE_REPLICATION_REQUEST, SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP,
+    SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP, SUFFIX_SOURCE_VERSION_ID, SUFFIX_TAGGING_TIMESTAMP, get_header,
     header_compat::{MINIO_ENCRYPTION_PREFIX, RUSTFS_ENCRYPTION_PREFIX},
-    insert_header_map,
+    insert_header_map, insert_str,
     metadata_compat::{MINIO_INTERNAL_PREFIX, RUSTFS_INTERNAL_PREFIX},
 };
 use rustfs_utils::http::{
@@ -422,6 +424,9 @@ pub fn get_complete_multipart_upload_opts_with_replication_authorization(
         preserve_etag,
         ..Default::default()
     };
+    if replication_request {
+        apply_replication_timestamps_from_headers(headers, &mut opts);
+    }
     apply_replica_status_from_headers(headers, &mut opts, replication_request_authorized);
 
     fill_conditional_writes_opts_from_header(headers, &mut opts)?;
@@ -479,6 +484,7 @@ pub fn put_opts_from_headers_with_replication_authorization(
         if let Some(crc) = get_header(headers, SUFFIX_REPLICATION_SSEC_CRC) {
             insert_header_map(&mut opts.user_defined, SUFFIX_REPLICATION_SSEC_CRC, crc.into_owned());
         }
+        apply_replication_timestamps_from_headers(headers, &mut opts);
     }
     Ok(opts)
 }
@@ -504,6 +510,47 @@ fn replication_source_mtime(headers: &HeaderMap<HeaderValue>) -> Option<time::Of
     }
 }
 
+/// Parses one replication LWW timestamp header. Invalid values are dropped
+/// with a warning (same tolerance as [`replication_source_mtime`]) so a
+/// malformed source header cannot wedge the replication queue.
+fn replication_timestamp_header(headers: &HeaderMap<HeaderValue>, suffix: &str) -> Option<time::OffsetDateTime> {
+    let value = get_header(headers, suffix)?;
+    let value = value.trim();
+    match time::OffsetDateTime::parse(value, &time::format_description::well_known::Rfc3339) {
+        Ok(timestamp) => Some(timestamp),
+        Err(err) => {
+            tracing::warn!("Invalid {} value '{}' (replication request=true): {}", suffix, value, err);
+            None
+        }
+    }
+}
+
+/// Callers must gate on an authorized replication request: these headers are
+/// trusted source-cluster state, not client input.
+fn apply_replication_timestamps_from_headers(headers: &HeaderMap<HeaderValue>, opts: &mut ObjectOptions) {
+    opts.replication_tagging_timestamp = replication_timestamp_header(headers, SUFFIX_SOURCE_REPLICATION_TAGGING_TIMESTAMP);
+    opts.replication_retention_timestamp = replication_timestamp_header(headers, SUFFIX_SOURCE_REPLICATION_RETENTION_TIMESTAMP);
+    opts.replication_legalhold_timestamp = replication_timestamp_header(headers, SUFFIX_SOURCE_REPLICATION_LEGALHOLD_TIMESTAMP);
+
+    // Persist into the internal metadata keys so a later outbound replication
+    // pass (replication_target_boundary) reads the source's modification
+    // times instead of falling back to mod_time.
+    // TODO(P1-6): receiver-side LWW is still missing — when the stored
+    // per-category timestamp is newer than the inbound one, the existing
+    // tags/retention/legal-hold should win instead of being overwritten.
+    for (timestamp, suffix) in [
+        (opts.replication_tagging_timestamp, SUFFIX_TAGGING_TIMESTAMP),
+        (opts.replication_retention_timestamp, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP),
+        (opts.replication_legalhold_timestamp, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP),
+    ] {
+        if let Some(timestamp) = timestamp
+            && let Ok(value) = timestamp.format(&time::format_description::well_known::Rfc3339)
+        {
+            insert_str(&mut opts.user_defined, suffix, value);
+        }
+    }
+}
+
 fn apply_replica_status_from_headers(headers: &HeaderMap<HeaderValue>, opts: &mut ObjectOptions, authorized: bool) {
     if !authorized {
         return;
@@ -663,6 +710,13 @@ fn is_reserved_user_metadata_key(key: &str) -> bool {
         || starts_with_ignore_ascii_case(key, MINIO_INTERNAL_PREFIX)
         || starts_with_ignore_ascii_case(key, RUSTFS_ENCRYPTION_PREFIX)
         || starts_with_ignore_ascii_case(key, MINIO_ENCRYPTION_PREFIX)
+        // Replication transport names (source-replication timestamps,
+        // source-mtime/-etag/-version-id, ...). A bare stored key with one of
+        // these names is forwarded verbatim by the outbound replication
+        // header builder on a server-authorized request, so the receiver
+        // would persist attacker-chosen values as trusted internal LWW state.
+        || starts_with_ignore_ascii_case(key, "x-rustfs-source-")
+        || starts_with_ignore_ascii_case(key, "x-minio-source-")
 }
 
 fn stored_user_metadata_key(key: &str) -> String {
@@ -1599,6 +1653,136 @@ mod tests {
         assert!(opts_invalid.mod_time.is_none());
     }
 
+    /// A client PUT must not materialize the replication transport names as
+    /// bare stored user-metadata keys: the outbound replication header
+    /// builder forwards user metadata verbatim on a server-authorized
+    /// request, so a bare `x-rustfs-source-replication-*-timestamp` key would
+    /// deliver an attacker-chosen value into the replica's trusted internal
+    /// LWW state (for a tagless object nothing later overwrites it).
+    #[test]
+    fn test_replication_transport_names_cannot_be_forged_via_user_metadata() {
+        let mut headers = HeaderMap::new();
+        for name in [
+            "x-amz-meta-x-rustfs-source-replication-tagging-timestamp",
+            "x-amz-meta-x-minio-source-replication-legalhold-timestamp",
+            "x-rustfs-meta-x-rustfs-source-replication-retention-timestamp",
+            "x-amz-meta-x-rustfs-source-mtime",
+        ] {
+            headers.insert(
+                http::header::HeaderName::from_static(name),
+                HeaderValue::from_static("2026-01-02T03:04:05Z"),
+            );
+        }
+
+        let metadata = extract_metadata(&headers);
+
+        for forged in [
+            "x-rustfs-source-replication-tagging-timestamp",
+            "x-minio-source-replication-legalhold-timestamp",
+            "x-rustfs-source-replication-retention-timestamp",
+            "x-rustfs-source-mtime",
+        ] {
+            assert!(
+                metadata.get(forged).is_none(),
+                "{forged} must not be storable as a bare user-metadata key"
+            );
+        }
+        // The values survive, namespaced back under the user-metadata prefix.
+        assert_eq!(
+            metadata
+                .get("x-amz-meta-x-rustfs-source-replication-tagging-timestamp")
+                .map(String::as_str),
+            Some("2026-01-02T03:04:05Z")
+        );
+    }
+
+    #[test]
+    fn test_put_opts_from_headers_gates_replication_timestamp_persistence_on_authorization() {
+        // Sender-side LWW state (replication_target_boundary.rs) is read back
+        // from these internal metadata keys, so an authorized replication PUT
+        // must persist the inbound timestamp headers; an unauthorized client
+        // must not be able to forge them.
+        let mut headers = HeaderMap::new();
+        insert_header(&mut headers, SUFFIX_SOURCE_REPLICATION_REQUEST, "true");
+        insert_header(&mut headers, "source-replication-tagging-timestamp", "2026-01-02T03:04:05Z");
+        insert_header(&mut headers, "source-replication-retention-timestamp", "2026-01-02T03:04:06Z");
+        insert_header(&mut headers, "source-replication-legalhold-timestamp", "2026-01-02T03:04:07Z");
+
+        let untrusted = put_opts_from_headers(&headers, HashMap::new()).expect("ordinary PUT options should be created");
+        for suffix in [
+            "tagging-timestamp",
+            "objectlock-retention-timestamp",
+            "objectlock-legalhold-timestamp",
+        ] {
+            assert!(
+                rustfs_utils::http::get_str(&untrusted.user_defined, suffix).is_none(),
+                "unauthorized clients must not persist the {suffix} internal key"
+            );
+        }
+        assert!(untrusted.replication_tagging_timestamp.is_none());
+        assert!(untrusted.replication_retention_timestamp.is_none());
+        assert!(untrusted.replication_legalhold_timestamp.is_none());
+
+        let trusted = put_opts_from_headers_with_replication_authorization(&headers, HashMap::new(), true)
+            .expect("authorized replication request should parse");
+        let parse = |value: &str| {
+            time::OffsetDateTime::parse(value, &time::format_description::well_known::Rfc3339).expect("valid RFC3339")
+        };
+        assert_eq!(trusted.replication_tagging_timestamp, Some(parse("2026-01-02T03:04:05Z")));
+        assert_eq!(trusted.replication_retention_timestamp, Some(parse("2026-01-02T03:04:06Z")));
+        assert_eq!(trusted.replication_legalhold_timestamp, Some(parse("2026-01-02T03:04:07Z")));
+        for (suffix, expected) in [
+            ("tagging-timestamp", "2026-01-02T03:04:05Z"),
+            ("objectlock-retention-timestamp", "2026-01-02T03:04:06Z"),
+            ("objectlock-legalhold-timestamp", "2026-01-02T03:04:07Z"),
+        ] {
+            assert_eq!(
+                rustfs_utils::http::get_str(&trusted.user_defined, suffix).as_deref(),
+                Some(expected),
+                "authorized replication must persist the {suffix} internal key"
+            );
+        }
+    }
+
+    #[test]
+    fn test_complete_multipart_opts_persist_replication_timestamps_when_authorized() {
+        let mut headers = HeaderMap::new();
+        insert_header(&mut headers, SUFFIX_SOURCE_REPLICATION_REQUEST, "true");
+        insert_header(&mut headers, "replication-actual-object-size", "1");
+        insert_header(&mut headers, "source-replication-tagging-timestamp", "2026-01-02T03:04:05Z");
+        insert_header(&mut headers, "source-replication-retention-timestamp", "2026-01-02T03:04:06Z");
+        insert_header(&mut headers, "source-replication-legalhold-timestamp", "2026-01-02T03:04:07Z");
+
+        let untrusted = get_complete_multipart_upload_opts(&headers).expect("ordinary multipart options should be created");
+        for suffix in [
+            "tagging-timestamp",
+            "objectlock-retention-timestamp",
+            "objectlock-legalhold-timestamp",
+        ] {
+            assert!(
+                rustfs_utils::http::get_str(&untrusted.user_defined, suffix).is_none(),
+                "unauthorized multipart completes must not persist the {suffix} internal key"
+            );
+        }
+
+        let trusted = get_complete_multipart_upload_opts_with_replication_authorization(&headers, true)
+            .expect("authorized multipart complete should parse");
+        for (suffix, expected) in [
+            ("tagging-timestamp", "2026-01-02T03:04:05Z"),
+            ("objectlock-retention-timestamp", "2026-01-02T03:04:06Z"),
+            ("objectlock-legalhold-timestamp", "2026-01-02T03:04:07Z"),
+        ] {
+            assert_eq!(
+                rustfs_utils::http::get_str(&trusted.user_defined, suffix).as_deref(),
+                Some(expected),
+                "authorized multipart completes must persist the {suffix} internal key"
+            );
+        }
+        assert!(trusted.replication_tagging_timestamp.is_some());
+        assert!(trusted.replication_retention_timestamp.is_some());
+        assert!(trusted.replication_legalhold_timestamp.is_some());
+    }
+
     #[test]
     fn test_put_opts_from_headers_with_replica_status() {
         let mut headers = HeaderMap::new();

scripts/security/check_preview_release_workflow.sh

@@ -195,6 +195,13 @@ IFS= read -r -d '' expected_docker_automatic_guard <<'EOF' || true
 EOF
 expected_docker_automatic_guard=${expected_docker_automatic_guard%$'\n'}
 require_job_if "$docker_workflow" "build-check" "$expected_docker_automatic_guard"
+require_line "$docker_workflow" '      source_ref: ${{ steps.check.outputs.source_ref }}' "Docker source ref output"
+require_line "$docker_workflow" '            source_ref="$HEAD_SHA"' "automatic Docker source ref"
+require_line "$docker_workflow" '              source_ref="$tag_ref"' "manual Docker source ref"
+require_line "$docker_workflow" '          ref: ${{ needs.build-check.outputs.source_ref }}' "Docker release source checkout"
+require_line "$docker_workflow" '          SOURCE_REVISION="$(git rev-parse HEAD)"' "Docker source revision resolution"
+require_line "$docker_workflow" '          LABELS="$LABELS,org.opencontainers.image.revision=$SOURCE_REVISION"' "Docker revision label"
+require_absent "$docker_workflow" 'org.opencontainers.image.revision=${{ github.sha }}' "Docker revision must not use the workflow branch SHA"
 
 docker_manual_guard=$(awk '
   $0 == "              *-preview*)" { in_preview = 1 }