fix(object-lock): require source timestamps before replication passes WORM gate

Adversarial review of the replication WORM bypass found a silent unlock:
an authorized replication write that carries no source timestamp for a
category (the source never held the object, only its tags changed) is
not judged by receiver-side LWW, so the metadata replace would drop the
destination's legal hold or retention unjudged. Before the bypass that
write was merely rejected.

Gate the bypass on `replication_write_may_pass_worm_gate`: the write
passes only when it carries the source timestamp of every category that
currently locks the version, so LWW decides each one; otherwise it stays
WORM-rejected. The set layer evaluates the lock gate first so malformed
persisted lock metadata still fails closed, and the app-layer pre-check
applies the same rule.

Refs rustfs/backlog#1953

crates/e2e_test/src/lib.rs

@@ -61,6 +61,11 @@ mod get_codec_streaming_compat_test;
 #[cfg(test)]
 mod version_id_regression_test;
 
+// Receiver-side replication LWW (rustfs/backlog#1953): stale inbound
+// replication metadata must not overwrite a newer local category state.
+#[cfg(test)]
+mod replication_lww_receiver_test;
+
 // Data usage regression tests
 #[cfg(test)]
 mod data_usage_test;

crates/e2e_test/src/replication_lww_receiver_test.rs

@@ -0,0 +1,152 @@
+#![cfg(test)]
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Receiver-side replication LWW over the wire (rustfs/backlog#1953, audit
+//! A4/P1-6).
+//!
+//! In an active-active topology both sites' metadata states arrive at the
+//! peer as authorized replication PUTs carrying per-category source
+//! timestamps (`x-rustfs-source-replication-tagging-timestamp` header
+//! family). Before the fix the receiver applied them unconditionally, so a
+//! stale delivery overwrote a newer local state and the two sites diverged
+//! permanently while both reported COMPLETED. This test drives one live
+//! `rustfs` server with simulated inbound replication PUTs for the same
+//! object version and asserts the newer tagging state wins regardless of
+//! delivery order, while a stale delivery still succeeds at the object level
+//! (a failure would loop through MRF re-delivering the stale value).
+
+use crate::common::{RustFSTestEnvironment, init_logging};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+use serial_test::serial;
+
+type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;
+
+const HDR_SOURCE_REPLICATION_REQUEST: &str = "x-rustfs-source-replication-request";
+const HDR_SOURCE_VERSION_ID: &str = "x-rustfs-source-version-id";
+const HDR_SOURCE_MTIME: &str = "x-rustfs-source-mtime";
+const HDR_SOURCE_TAGGING_TIMESTAMP: &str = "x-rustfs-source-replication-tagging-timestamp";
+
+const SOURCE_MTIME: &str = "2026-01-01T00:00:00Z";
+const T_STALE: &str = "2026-01-01T00:00:01Z";
+const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+const T_NEWER: &str = "2026-03-01T00:00:00Z";
+
+/// Simulated inbound authorized replication PUT: same object version, tags and
+/// the source-authored tagging timestamp carried in transport headers.
+async fn inbound_replication_put(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    tags: &str,
+    tagging_timestamp: &str,
+) -> TestResult {
+    let version_id = version_id.to_string();
+    let tagging_timestamp = tagging_timestamp.to_string();
+    client
+        .put_object()
+        .bucket(bucket)
+        .key(key)
+        .body(ByteStream::from_static(b"lww-e2e-body"))
+        .tagging(tags)
+        .customize()
+        .mutate_request(move |req| {
+            req.headers_mut().insert(HDR_SOURCE_REPLICATION_REQUEST, "true");
+            req.headers_mut().insert(HDR_SOURCE_VERSION_ID, version_id.clone());
+            req.headers_mut().insert(HDR_SOURCE_MTIME, SOURCE_MTIME);
+            req.headers_mut()
+                .insert(HDR_SOURCE_TAGGING_TIMESTAMP, tagging_timestamp.clone());
+        })
+        .send()
+        .await?;
+    Ok(())
+}
+
+async fn tag_value(client: &Client, bucket: &str, key: &str, version_id: &str, tag_key: &str) -> Option<String> {
+    let tagging = client
+        .get_object_tagging()
+        .bucket(bucket)
+        .key(key)
+        .version_id(version_id)
+        .send()
+        .await
+        .expect("object tagging should be readable");
+    tagging
+        .tag_set()
+        .iter()
+        .find(|tag| tag.key() == tag_key)
+        .map(|tag| tag.value().to_string())
+}
+
+#[tokio::test(flavor = "multi_thread")]
+#[serial]
+async fn receiver_lww_keeps_newer_tags_across_delivery_orders() -> TestResult {
+    init_logging();
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server(vec![]).await?;
+    let client = env.create_s3_client();
+
+    let bucket = "replication-lww-receiver";
+    let key = "object";
+    client.create_bucket().bucket(bucket).send().await?;
+    client
+        .put_bucket_versioning()
+        .bucket(bucket)
+        .versioning_configuration(
+            VersioningConfiguration::builder()
+                .status(BucketVersioningStatus::Enabled)
+                .build(),
+        )
+        .send()
+        .await?;
+
+    // First delivery establishes version V with tags stamped T_LOCAL.
+    let version_id = uuid::Uuid::new_v4().to_string();
+    inbound_replication_put(&client, bucket, key, &version_id, "site=local", T_LOCAL).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("local"),
+        "the first delivery must establish the tagged version"
+    );
+
+    // A stale delivery (older source timestamp) must succeed at the object
+    // level but must NOT overwrite the newer tags.
+    inbound_replication_put(&client, bucket, key, &version_id, "site=stale", T_STALE).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("local"),
+        "a stale inbound delivery must not overwrite newer tags (rustfs/backlog#1953)"
+    );
+
+    // A newer delivery still converges the version onto the newest state.
+    inbound_replication_put(&client, bucket, key, &version_id, "site=newer", T_NEWER).await?;
+    assert_eq!(
+        tag_value(&client, bucket, key, &version_id, "site").await.as_deref(),
+        Some("newer"),
+        "a newer inbound delivery must overwrite older tags"
+    );
+
+    client
+        .delete_object()
+        .bucket(bucket)
+        .key(key)
+        .version_id(&version_id)
+        .send()
+        .await?;
+    env.delete_test_bucket(bucket).await.ok();
+    Ok(())
+}

crates/ecstore/src/api/mod.rs

@@ -161,7 +161,7 @@ pub mod bucket {
         pub mod objectlock_sys {
             pub use crate::bucket::object_lock::objectlock_sys::{
                 BucketObjectLockSys, ObjectLockBlockReason, add_years, check_object_lock_for_deletion,
-                check_retention_for_modification, is_retention_active,
+                check_retention_for_modification, is_retention_active, replication_write_may_pass_worm_gate,
             };
         }
     }

crates/ecstore/src/bucket/object_lock/objectlock_sys.rs

@@ -15,7 +15,7 @@
 use crate::bucket::metadata_sys::{ObjectLockConfigState, get_object_lock_config, get_object_lock_config_state};
 use crate::bucket::object_lock::objectlock;
 use crate::error::{Error, Result, StorageError};
-use crate::object_api::ObjectInfo;
+use crate::object_api::{ObjectInfo, ObjectOptions};
 use s3s::dto::{Date, DefaultRetention, ObjectLockConfiguration, ObjectLockLegalHoldStatus, ObjectLockRetentionMode};
 use s3s::header::{X_AMZ_OBJECT_LOCK_LEGAL_HOLD, X_AMZ_OBJECT_LOCK_MODE, X_AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE};
 use std::sync::Arc;
@@ -136,12 +136,41 @@ pub fn add_years(dt: OffsetDateTime, years: i32) -> OffsetDateTime {
 
 /// Check if an object has legal hold enabled.
 /// Returns true if legal hold is ON.
-#[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 fn has_legal_hold(user_defined: &std::collections::HashMap<String, String>) -> bool {
     let lhold = objectlock::get_object_legalhold_meta(user_defined);
     matches!(lhold.status, Some(ref st) if st.as_str() == ObjectLockLegalHoldStatus::ON)
 }
 
+/// Whether an authorized replication write (`ObjectOptions::replication_request`)
+/// may overwrite a locked destination version.
+///
+/// The source's lock state governs a replica (MinIO `checkPutObjectLockAllowed`
+/// skips the existing-version check for replicas), and a source-side hold
+/// release or retention change reaches this site only through this write. The
+/// overwrite is allowed only when the write carries the source timestamp of
+/// every category that currently locks the version, so receiver-side LWW
+/// (`merge_replication_metadata_lww`) judges each of them: a category locked
+/// more recently here is kept, otherwise the source's newer state wins. A write
+/// without that timestamp carries no source decision for the category — the
+/// metadata replace would lift the lock unjudged — so it stays WORM-rejected.
+pub fn replication_write_may_pass_worm_gate(
+    user_defined: &std::collections::HashMap<String, String>,
+    opts: &ObjectOptions,
+) -> bool {
+    if !opts.replication_request {
+        return false;
+    }
+    if has_legal_hold(user_defined) && opts.replication_legalhold_timestamp.is_none() {
+        return false;
+    }
+    let ret = objectlock::get_object_retention_meta(user_defined);
+    let retention_locked = ret
+        .mode
+        .as_ref()
+        .is_some_and(|mode| is_retention_active(mode.as_str(), ret.retain_until_date.as_ref()));
+    !(retention_locked && opts.replication_retention_timestamp.is_none())
+}
+
 /// Check if an object is locked based on its metadata.
 /// This is a common function used by both lifecycle evaluation and deletion checks.
 ///
@@ -239,7 +268,12 @@ pub(crate) fn check_object_lock_for_deletion_with_config(
         return Ok(None);
     }
 
-    if let Some(status) = obj_info.user_defined.get(X_AMZ_OBJECT_LOCK_LEGAL_HOLD.as_str()) {
+    // A cleared retention / legal hold is persisted as empty strings (the
+    // MinIO on-disk shape, `parse_object_lock_retention`); read it as "no lock"
+    // rather than as corrupt metadata.
+    let persisted = |key: &str| obj_info.user_defined.get(key).filter(|value| !value.is_empty());
+
+    if let Some(status) = persisted(X_AMZ_OBJECT_LOCK_LEGAL_HOLD.as_str()) {
         if status.eq_ignore_ascii_case(ObjectLockLegalHoldStatus::ON) {
             return Ok(Some(ObjectLockBlockReason::LegalHold));
         }
@@ -248,8 +282,8 @@ pub(crate) fn check_object_lock_for_deletion_with_config(
         }
     }
 
-    let mode = obj_info.user_defined.get(X_AMZ_OBJECT_LOCK_MODE.as_str());
-    let retain_until = obj_info.user_defined.get(X_AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE.as_str());
+    let mode = persisted(X_AMZ_OBJECT_LOCK_MODE.as_str());
+    let retain_until = persisted(X_AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE.as_str());
     let explicit_ret = match (mode, retain_until) {
         (None, None) => None,
         (Some(mode), Some(retain_until)) => {
@@ -486,6 +520,98 @@ mod tests {
         }
     }
 
+    /// A replication write passes the WORM gate only when it carries the
+    /// source timestamp of every category that currently locks the version.
+    #[test]
+    fn replication_write_passes_worm_gate_only_with_every_locking_category_timestamp() {
+        use rustfs_utils::http::headers::{
+            AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER,
+        };
+
+        let hold = [(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, "ON")];
+        let retention = [
+            (AMZ_OBJECT_LOCK_MODE_LOWER, "GOVERNANCE"),
+            (AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, "2099-01-01T00:00:00Z"),
+        ];
+        let expired = [
+            (AMZ_OBJECT_LOCK_MODE_LOWER, "COMPLIANCE"),
+            (AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, "2000-01-01T00:00:00Z"),
+        ];
+        let metadata = |entries: &[&[(&str, &str)]]| -> std::collections::HashMap<String, String> {
+            entries
+                .iter()
+                .flat_map(|entries| entries.iter())
+                .map(|(key, value)| (key.to_string(), value.to_string()))
+                .collect()
+        };
+        let opts = |hold_ts: bool, retention_ts: bool| ObjectOptions {
+            replication_request: true,
+            replication_legalhold_timestamp: hold_ts.then_some(OffsetDateTime::UNIX_EPOCH),
+            replication_retention_timestamp: retention_ts.then_some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        };
+
+        let locked_by_both = metadata(&[&hold, &retention]);
+        assert!(replication_write_may_pass_worm_gate(&locked_by_both, &opts(true, true)));
+        assert!(!replication_write_may_pass_worm_gate(&locked_by_both, &opts(true, false)));
+        assert!(!replication_write_may_pass_worm_gate(&locked_by_both, &opts(false, true)));
+
+        assert!(replication_write_may_pass_worm_gate(&metadata(&[&hold]), &opts(true, false)));
+        assert!(!replication_write_may_pass_worm_gate(&metadata(&[&hold]), &opts(false, true)));
+        assert!(replication_write_may_pass_worm_gate(&metadata(&[&retention]), &opts(false, true)));
+        assert!(!replication_write_may_pass_worm_gate(&metadata(&[&retention]), &opts(true, false)));
+
+        // Expired retention and a released hold no longer lock anything.
+        let unlocked = metadata(&[&expired, &[(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, "OFF")]]);
+        assert!(replication_write_may_pass_worm_gate(&unlocked, &opts(false, false)));
+
+        // Never for a non-replication write, whatever it carries.
+        let local = ObjectOptions {
+            replication_request: false,
+            ..opts(true, true)
+        };
+        assert!(!replication_write_may_pass_worm_gate(&metadata(&[&hold]), &local));
+    }
+
+    /// A local PutObjectRetention / PutObjectLegalHold "clear" persists the
+    /// lock keys as empty strings (the MinIO on-disk shape, see
+    /// `parse_object_lock_retention`); that is "no lock", not corruption, and
+    /// must not wedge later explicit-version PUTs or deletes
+    /// (rustfs/backlog#1953).
+    #[test]
+    fn deletion_treats_cleared_empty_lock_metadata_as_unlocked() {
+        use rustfs_utils::http::headers::{
+            AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER,
+        };
+
+        let cases: [(&str, &[&str]); 3] = [
+            (
+                "cleared retention",
+                &[AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER],
+            ),
+            ("cleared legal hold", &[AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER]),
+            (
+                "all cleared",
+                &[
+                    AMZ_OBJECT_LOCK_MODE_LOWER,
+                    AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER,
+                    AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER,
+                ],
+            ),
+        ];
+
+        for (case, keys) in cases {
+            let user_defined = keys.iter().map(|key| (key.to_string(), String::new())).collect();
+            let obj_info = ObjectInfo {
+                user_defined: Arc::new(user_defined),
+                ..Default::default()
+            };
+
+            let result = check_object_lock_for_deletion_with_config(None, &obj_info, false);
+            assert!(matches!(result, Ok(None)), "{case}: empty lock keys must read as unlocked: {result:?}");
+        }
+    }
+
     #[test]
     fn deletion_rejects_invalid_persisted_legal_hold_metadata() {
         let mut user_defined = std::collections::HashMap::new();

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

@@ -3868,6 +3868,7 @@ async fn replicate_object_with_multipart<S: ReplicationObjectIO>(ctx: MultipartR
                 actual_size,
                 object_info.etag.clone().unwrap_or_default(),
                 object_info.mod_time,
+                &put_opts.internal,
             ),
         )
         .await

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

@@ -472,6 +472,7 @@ pub(crate) fn replication_complete_multipart_options(
     actual_size: String,
     source_etag: String,
     source_mtime: Option<OffsetDateTime>,
+    source_internal: &AdvancedPutOptions,
 ) -> PutObjectOptions {
     let mut user_metadata = HashMap::new();
     insert_header_map(&mut user_metadata, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE, actual_size);
@@ -484,6 +485,14 @@ pub(crate) fn replication_complete_multipart_options(
             // mtime must degrade to epoch so header() suppresses the header
             // instead of asserting the replication time as the object's mtime.
             source_mtime: source_mtime.unwrap_or(OffsetDateTime::UNIX_EPOCH),
+            // Carry the per-category LWW timestamps on the complete request as
+            // well: the receiver's CompleteMultipartUpload options builder
+            // parses the same headers, so the multipart transport gets the
+            // same receiver-side LWW as the single-PUT transport
+            // (rustfs/backlog#1953). Epoch values keep the headers suppressed.
+            tagging_timestamp: source_internal.tagging_timestamp,
+            retention_timestamp: source_internal.retention_timestamp,
+            legalhold_timestamp: source_internal.legalhold_timestamp,
             replication_status: ReplicationStatusType::Replica,
             replication_request: true,
             ..Default::default()
@@ -649,20 +658,39 @@ mod tests {
     #[test]
     fn replication_complete_multipart_options_sets_actual_size() {
         let source_mtime = OffsetDateTime::from_unix_timestamp(1_716_170_000).expect("valid test timestamp");
+        let source_internal = AdvancedPutOptions {
+            tagging_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_100).expect("valid test timestamp"),
+            retention_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_200).expect("valid test timestamp"),
+            legalhold_timestamp: OffsetDateTime::from_unix_timestamp(1_716_170_300).expect("valid test timestamp"),
+            ..Default::default()
+        };
         let options = replication_complete_multipart_options(
             "1024".to_string(),
             "0123456789abcdef0123456789abcdef-3".to_string(),
             Some(source_mtime),
+            &source_internal,
         );
         assert_eq!(options.internal.source_etag, "0123456789abcdef0123456789abcdef-3");
         assert_eq!(options.internal.source_mtime, source_mtime);
 
+        // The complete request must carry the same per-category LWW timestamps
+        // as the initiate request; the receiver reads them from the complete
+        // headers (rustfs/backlog#1953).
+        assert_eq!(options.internal.tagging_timestamp, source_internal.tagging_timestamp);
+        assert_eq!(options.internal.retention_timestamp, source_internal.retention_timestamp);
+        assert_eq!(options.internal.legalhold_timestamp, source_internal.legalhold_timestamp);
+
         // Absent source mtime must degrade to epoch (header suppressed), not
         // the AdvancedPutOptions default of now_utc() — that default would
         // stamp the replication time as the replica's mtime and break the
-        // multipart HEAD convergence.
-        let options_no_mtime = replication_complete_multipart_options("1024".to_string(), String::new(), None);
+        // multipart HEAD convergence. Unset category timestamps stay epoch so
+        // header() keeps suppressing them.
+        let options_no_mtime =
+            replication_complete_multipart_options("1024".to_string(), String::new(), None, &AdvancedPutOptions::default());
         assert_eq!(options_no_mtime.internal.source_mtime.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.tagging_timestamp.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.retention_timestamp.unix_timestamp(), 0);
+        assert_eq!(options_no_mtime.internal.legalhold_timestamp.unix_timestamp(), 0);
 
         assert_eq!(
             get_header_map(&options.user_metadata, SUFFIX_REPLICATION_ACTUAL_OBJECT_SIZE).as_deref(),

crates/ecstore/src/set_disk/mod.rs

@@ -47,6 +47,7 @@ use crate::bucket::metadata_sys;
 use crate::bucket::metadata_sys::ObjectLockConfigState;
 use crate::bucket::object_lock::objectlock_sys::{
     check_object_lock_for_deletion_with_config, check_object_lock_for_deletion_with_state, check_retention_for_modification,
+    replication_write_may_pass_worm_gate,
 };
 use crate::bucket::replication::{
     ReplicateDecision, ReplicationObjectBridge, ReplicationState, ReplicationStatusType, VersionPurgeStatusType,

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

@@ -2223,6 +2223,57 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
             fi.set_data_moved();
         }
 
+        // Receiver-side LWW (rustfs/backlog#1953): the multipart replication
+        // transport carries the category values at CreateMultipartUpload (in
+        // the staged upload metadata) and the source category timestamps on
+        // the complete request. Read the destination version under the held
+        // object write lock and keep any category this site modified more
+        // recently. Read failures (version absent on first replication, quorum
+        // errors) keep today's overwrite semantics: failing the complete would
+        // loop through MRF, re-delivering the stale value forever.
+        if crate::set_disk::ops::object::replication_lww_applicable(opts)
+            && let Some(version_id) = fi.version_id
+        {
+            match self
+                .get_object_info(
+                    bucket,
+                    object,
+                    &ObjectOptions {
+                        version_id: Some(version_id.to_string()),
+                        no_lock: true,
+                        metadata_cache_safe: false,
+                        versioned: opts.versioned,
+                        version_suspended: opts.version_suspended,
+                        ..Default::default()
+                    },
+                )
+                .await
+            {
+                Ok(existing) => {
+                    let stored = crate::set_disk::ops::object::stored_replication_category_metadata(&existing);
+                    crate::set_disk::ops::object::merge_replication_metadata_lww(&mut fi.metadata, &stored, opts);
+                }
+                // Version absent: first replication of this version, nothing
+                // local to compare — the normal path, not a degraded one.
+                Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
+                Err(err) => {
+                    // Degraded path: without the stored state the inbound
+                    // metadata is applied unchanged — exactly the overwrite
+                    // LWW exists to prevent — so this must be operator-visible.
+                    warn!(
+                        component = LOG_COMPONENT_ECSTORE,
+                        subsystem = LOG_SUBSYSTEM_SET_DISK,
+                        bucket,
+                        object,
+                        version_id = %version_id,
+                        error = %err,
+                        state = "replication_lww_read_unavailable",
+                        "SetDisk multipart replication LWW read skipped; inbound metadata applied without comparison"
+                    );
+                }
+            }
+        }
+
         for meta in parts_metadatas.iter_mut() {
             if meta.has_valid_erasure_geometry() {
                 meta.size = fi.size;
@@ -6960,6 +7011,97 @@ mod tests {
                 .await
         }
 
+        /// Receiver-side LWW on the multipart replication transport
+        /// (rustfs/backlog#1953): a metadata-only replication of a multipart
+        /// source object rides CreateMultipartUpload (category values in the
+        /// upload metadata) + CompleteMultipartUpload (category timestamps in
+        /// the complete options). A stale inbound tagging timestamp must not
+        /// overwrite a newer locally-tagged destination version.
+        #[tokio::test]
+        #[serial]
+        async fn complete_multipart_upload_stale_replication_tags_keep_local() {
+            use rustfs_utils::http::headers::AMZ_OBJECT_TAGGING;
+            use rustfs_utils::http::{SUFFIX_TAGGING_TIMESTAMP, get_str};
+            use time::format_description::well_known::Rfc3339;
+
+            const T_OLD: &str = "2026-01-01T00:00:00Z";
+            const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+
+            let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+            let bucket = "multipart-replication-lww-bucket";
+            let object = "object";
+            make_bucket_on_all(&disk_stores, bucket).await;
+
+            // Local destination version with newer tags.
+            let version_id = Uuid::new_v4();
+            let mut local_metadata = HashMap::new();
+            local_metadata.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+            rustfs_utils::http::insert_str(&mut local_metadata, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+            let mut local_reader = PutObjReader::from_vec(b"local body".to_vec());
+            set_disks
+                .put_object(
+                    bucket,
+                    object,
+                    &mut local_reader,
+                    &ObjectOptions {
+                        versioned: true,
+                        version_id: Some(version_id.to_string()),
+                        user_defined: local_metadata,
+                        // Explicit-version PUTs require the bucket Object Lock snapshot.
+                        object_lock_config_snapshot: Some(Arc::new(crate::set_disk::ObjectLockConfigSnapshot::new(
+                            crate::bucket::metadata_sys::ObjectLockConfigState::ConfirmedAbsent,
+                        ))),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("local versioned put should commit");
+
+            // Inbound replication upload carrying older tags for the same version.
+            let mut inbound_metadata = HashMap::new();
+            inbound_metadata.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+            rustfs_utils::http::insert_str(&mut inbound_metadata, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+            let create_opts = ObjectOptions {
+                versioned: true,
+                user_defined: inbound_metadata,
+                ..Default::default()
+            };
+            let (upload_id, parts) =
+                stage_upload_with_create_opts(&set_disks, bucket, object, &payload(0x5a), &create_opts).await;
+            rewrite_staged_upload_version_id(&set_disks, bucket, object, &upload_id, Some(version_id)).await;
+
+            let complete_opts = ObjectOptions {
+                versioned: true,
+                replication_request: true,
+                replication_tagging_timestamp: Some(OffsetDateTime::parse(T_OLD, &Rfc3339).expect("test timestamp should parse")),
+                ..Default::default()
+            };
+            set_disks
+                .clone()
+                .complete_multipart_upload(bucket, object, &upload_id, parts, &complete_opts)
+                .await
+                .expect("replication multipart completion should succeed even when a category keeps local values");
+
+            let info = set_disks
+                .get_object_info(
+                    bucket,
+                    object,
+                    &ObjectOptions {
+                        versioned: true,
+                        version_id: Some(version_id.to_string()),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("completed version should be readable");
+            assert_eq!(
+                info.user_tags.as_str(),
+                "site=local",
+                "older inbound multipart tags must not overwrite newer local tags"
+            );
+            assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_LOCAL));
+        }
+
         #[tokio::test]
         #[serial]
         async fn complete_multipart_upload_assigns_completion_version_id() {

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

@@ -1880,6 +1880,110 @@ fn delete_file_info_with_replication_transport_metadata(fi: &FileInfo) -> FileIn
     transported
 }
 
+/// True when an authorized replication write carries at least one per-category
+/// source timestamp, i.e. receiver-side LWW has something to judge.
+pub(in crate::set_disk) fn replication_lww_applicable(opts: &ObjectOptions) -> bool {
+    opts.replication_request
+        && (opts.replication_tagging_timestamp.is_some()
+            || opts.replication_retention_timestamp.is_some()
+            || opts.replication_legalhold_timestamp.is_some())
+}
+
+/// The stored per-category state of a destination version, as compared by
+/// [`merge_replication_metadata_lww`]. `ObjectInfo::from_file_info`
+/// externalizes tags into `user_tags` (stripping the metadata key), so the
+/// tag value is folded back into map form here.
+pub(in crate::set_disk) fn stored_replication_category_metadata(existing: &ObjectInfo) -> HashMap<String, String> {
+    let mut stored = (*existing.user_defined).clone();
+    if !existing.user_tags.is_empty() {
+        stored.insert(rustfs_utils::http::headers::AMZ_OBJECT_TAGGING.to_string(), (*existing.user_tags).clone());
+    }
+    stored
+}
+
+/// Receiver-side last-writer-wins for authorized replication writes
+/// (rustfs/backlog#1953, audit A4/P1-6). Metadata-only replication reuses the
+/// whole-object transports, so in active-active topologies an inbound write
+/// carries the source's tags / retention / legal hold verbatim and would
+/// otherwise overwrite a category the destination modified more recently —
+/// both sites end up permanently diverged while reporting COMPLETED.
+///
+/// Judged per category, only when the inbound request carries that category's
+/// source timestamp (`ObjectOptions::replication_*_timestamp`):
+/// - stored timestamp newer than inbound: the local category values and
+///   timestamp are kept; the rest of the write proceeds per the inbound
+///   metadata and the object-level result stays successful (failing the write
+///   instead would loop through MRF, re-delivering the stale value forever);
+/// - otherwise the inbound category wins and its internal timestamp key is
+///   pinned to the source-authored time — the PUT path re-stamps the
+///   object-lock timestamps with the receiver's clock
+///   (`parse_object_lock_retention` / `parse_object_lock_legal_hold` insert
+///   `now()` via `eval_metadata`), which would make the replica's clock the
+///   LWW authority and wedge later convergence;
+/// - no stored timestamp (pre-P1-6 data) or no inbound timestamp: the current
+///   overwrite behavior is preserved.
+///
+/// Returns whether `inbound` was modified. Callers must hold the object write
+/// lock so the stored values compared here are the ones being replaced.
+pub(in crate::set_disk) fn merge_replication_metadata_lww(
+    inbound: &mut HashMap<String, String>,
+    existing: &HashMap<String, String>,
+    opts: &ObjectOptions,
+) -> bool {
+    use rustfs_utils::http::headers::{
+        AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, AMZ_OBJECT_TAGGING,
+    };
+    use rustfs_utils::http::metadata_compat::{
+        SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, SUFFIX_TAGGING_TIMESTAMP, get_str,
+        remove_str,
+    };
+    use time::format_description::well_known::Rfc3339;
+
+    let categories: [(Option<OffsetDateTime>, &str, &[&str]); 3] = [
+        (opts.replication_tagging_timestamp, SUFFIX_TAGGING_TIMESTAMP, &[AMZ_OBJECT_TAGGING]),
+        (
+            opts.replication_retention_timestamp,
+            SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP,
+            &[AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER],
+        ),
+        (
+            opts.replication_legalhold_timestamp,
+            SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP,
+            &[AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER],
+        ),
+    ];
+
+    let mut changed = false;
+    for (inbound_timestamp, timestamp_suffix, value_keys) in categories {
+        let Some(inbound_timestamp) = inbound_timestamp else { continue };
+        let is_category_value_key = |key: &str| value_keys.iter().any(|value_key| key.eq_ignore_ascii_case(value_key));
+        let stored_timestamp = get_str(existing, timestamp_suffix).and_then(|value| OffsetDateTime::parse(&value, &Rfc3339).ok());
+        if stored_timestamp.is_some_and(|stored| stored > inbound_timestamp) {
+            inbound.retain(|key, _| !is_category_value_key(key));
+            remove_str(inbound, timestamp_suffix);
+            for (key, value) in existing {
+                if is_category_value_key(key) {
+                    inbound.insert(key.clone(), value.clone());
+                }
+            }
+            // Restore the winning timestamp via insert_str, not a verbatim key
+            // copy: a MinIO-written version may carry only the
+            // x-minio-internal- key, and the dual-key invariant requires every
+            // write to produce both keys.
+            if let Some(stored_value) = get_str(existing, timestamp_suffix) {
+                rustfs_utils::http::insert_str(inbound, timestamp_suffix, stored_value);
+            }
+            changed = true;
+        } else if let Ok(source_authored) = inbound_timestamp.format(&Rfc3339)
+            && get_str(inbound, timestamp_suffix).as_deref() != Some(source_authored.as_str())
+        {
+            rustfs_utils::http::insert_str(inbound, timestamp_suffix, source_authored);
+            changed = true;
+        }
+    }
+    changed
+}
+
 impl SetDisks {
     pub(in crate::set_disk) async fn persist_old_data_cleanup_receipts(
         &self,
@@ -2557,9 +2661,33 @@ impl SetDisks {
                         let object_lock_config = opts.object_lock_config_snapshot.as_deref().ok_or_else(|| {
                             Error::other("explicit-version PUT is missing its Object Lock configuration snapshot")
                         })?;
-                        if check_object_lock_for_deletion_with_state(object_lock_config.state(), &existing, false)?.is_some() {
+                        // The WORM gate protects the locked version from local
+                        // overwrites; an authorized replication write passes it
+                        // only when the LWW merge below will judge every
+                        // locking category (see
+                        // `replication_write_may_pass_worm_gate`). Gate first so
+                        // malformed lock metadata still fails closed.
+                        if check_object_lock_for_deletion_with_state(object_lock_config.state(), &existing, false)?.is_some()
+                            && !replication_write_may_pass_worm_gate(&existing.user_defined, opts)
+                        {
                             return Err(StorageError::PrefixAccessDenied(bucket.to_string(), object.to_string()));
                         }
+                        // Receiver-side LWW (rustfs/backlog#1953): reuse this
+                        // commit-lock read of the destination version so a
+                        // category (tags / retention / legal hold) modified
+                        // more recently on this site is kept instead of being
+                        // overwritten by the inbound replication metadata.
+                        if replication_lww_applicable(opts) {
+                            let stored = stored_replication_category_metadata(&existing);
+                            let mut merged = parts_metadatas[response_metadata_slot].metadata.clone();
+                            if merge_replication_metadata_lww(&mut merged, &stored, opts) {
+                                for (pfi, disk) in parts_metadatas.iter_mut().zip(shuffle_disks.iter()) {
+                                    if disk.is_some() {
+                                        pfi.metadata = merged.clone();
+                                    }
+                                }
+                            }
+                        }
                     }
                     Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
                     Err(err) => return Err(err),
@@ -7887,6 +8015,540 @@ mod replication_quota_safety_tests {
     }
 }
 
+#[cfg(test)]
+mod replication_lww_tests {
+    //! Receiver-side LWW for authorized replication writes (rustfs/backlog#1953,
+    //! audit A4/P1-6): an inbound replication PUT whose per-category timestamp
+    //! (tags / retention / legal hold) is older than the destination version's
+    //! stored timestamp must keep the local category values instead of
+    //! overwriting them; categories are judged independently and the write
+    //! itself still succeeds.
+
+    use super::hermetic_set_disks_support::hermetic_set_disks_isolated as hermetic_set_disks;
+    use super::*;
+    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
+    use rustfs_utils::http::headers::{
+        AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER, AMZ_OBJECT_LOCK_MODE_LOWER, AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER, AMZ_OBJECT_TAGGING,
+    };
+    use rustfs_utils::http::{
+        SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, SUFFIX_TAGGING_TIMESTAMP, get_str,
+        insert_str,
+    };
+    use time::format_description::well_known::Rfc3339;
+
+    const T_OLD: &str = "2026-01-01T00:00:00Z";
+    const T_LOCAL: &str = "2026-02-01T00:00:00Z";
+    const T_NEW: &str = "2026-03-01T00:00:00Z";
+
+    fn parse_ts(value: &str) -> OffsetDateTime {
+        OffsetDateTime::parse(value, &Rfc3339).expect("test timestamp should parse")
+    }
+
+    async fn make_bucket(disks: &[DiskStore], bucket: &str) {
+        for disk in disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+    }
+
+    async fn put_version(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str, opts: &ObjectOptions) {
+        let mut reader = PutObjReader::from_vec(b"lww-body".to_vec());
+        set_disks
+            .put_object(bucket, object, &mut reader, opts)
+            .await
+            .expect("versioned put should commit");
+        assert_eq!(opts.version_id.as_deref(), Some(version_id));
+    }
+
+    fn versioned_opts(version_id: &str, user_defined: HashMap<String, String>) -> ObjectOptions {
+        ObjectOptions {
+            versioned: true,
+            version_id: Some(version_id.to_string()),
+            user_defined,
+            // Explicit-version PUTs require the bucket Object Lock snapshot.
+            object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(
+                crate::bucket::metadata_sys::ObjectLockConfigState::ConfirmedAbsent,
+            ))),
+            ..Default::default()
+        }
+    }
+
+    /// Local state: version `version_id` with tags "site=local" stamped `T_LOCAL`.
+    async fn seed_local_tagged_version(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str) {
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        insert_str(&mut user_defined, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        put_version(set_disks, bucket, object, version_id, &versioned_opts(version_id, user_defined)).await;
+    }
+
+    fn inbound_tagging_opts(version_id: &str, tags: &str, timestamp: &str) -> ObjectOptions {
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), tags.to_string());
+        insert_str(&mut user_defined, SUFFIX_TAGGING_TIMESTAMP, timestamp.to_string());
+        ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(timestamp)),
+            ..versioned_opts(version_id, user_defined)
+        }
+    }
+
+    async fn version_info(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str) -> ObjectInfo {
+        set_disks
+            .get_object_info(bucket, object, &versioned_opts(version_id, HashMap::new()))
+            .await
+            .expect("version should be readable")
+    }
+
+    #[tokio::test]
+    async fn inbound_stale_tagging_keeps_newer_local_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-stale";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_local_tagged_version(&set_disks, bucket, object, &version_id).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=local",
+            "older inbound tags must not overwrite newer local tags"
+        );
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(),
+            Some(T_LOCAL),
+            "the winning local tagging timestamp must be preserved"
+        );
+    }
+
+    #[tokio::test]
+    async fn inbound_newer_tagging_overwrites_local_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-newer";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_local_tagged_version(&set_disks, bucket, object, &version_id).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_NEW),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=remote",
+            "newer inbound tags must overwrite older local tags"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_NEW));
+    }
+
+    #[tokio::test]
+    async fn inbound_wins_when_local_has_no_tagging_timestamp() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-no-local-ts";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        // Pre-P1-6 data: local tags without a stored tagging timestamp.
+        let mut user_defined = HashMap::new();
+        user_defined.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, user_defined)).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_tags.as_str(),
+            "site=remote",
+            "without a local timestamp the inbound category must win (pre-LWW data compatibility)"
+        );
+    }
+
+    #[tokio::test]
+    async fn categories_are_judged_independently() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-category-independent";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local: newer tags (T_LOCAL), older *cleared* retention (T_OLD) —
+        // timestamp key only, the shape a replicated retention clear stores.
+        // (An active local retention would already block the overwrite at the
+        // WORM gate; the LWW-reachable retention states are cleared/expired.)
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string());
+        insert_str(&mut local, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_OLD.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        // Inbound: older tags (T_OLD), newer retention (T_NEW).
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+        insert_str(&mut inbound, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "COMPLIANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2028-01-01T00:00:00Z".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_NEW.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(T_OLD)),
+            replication_retention_timestamp: Some(parse_ts(T_NEW)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(info.user_tags.as_str(), "site=local", "the stale tagging category must keep local values");
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_MODE_LOWER).map(String::as_str),
+            Some("COMPLIANCE"),
+            "the newer retention category must be applied in the same write"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP).as_deref(), Some(T_NEW));
+    }
+
+    #[tokio::test]
+    async fn inbound_stale_legal_hold_keeps_local_value() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-legalhold-stale";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local: legal hold released (OFF) at T_LOCAL. (A local hold that is
+        // still ON already blocks the overwrite at the WORM gate; the
+        // LWW-reachable divergence is a stale inbound ON resurrecting a hold
+        // that was released more recently on this site.)
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "OFF".to_string());
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, T_LOCAL.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "ON".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, T_OLD.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_legalhold_timestamp: Some(parse_ts(T_OLD)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str),
+            Some("OFF"),
+            "a stale inbound legal hold must not resurrect a hold released more recently"
+        );
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP).as_deref(),
+            Some(T_LOCAL)
+        );
+    }
+
+    /// Dual-key invariant under LWW: a MinIO-written destination version may
+    /// carry only the x-minio-internal timestamp key; when the local category
+    /// wins, the restored map must still hold BOTH compatibility keys.
+    #[test]
+    fn local_win_restores_both_internal_timestamp_keys_for_minio_only_metadata() {
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_TAGGING.to_string(), "site=remote".to_string());
+        insert_str(&mut inbound, SUFFIX_TAGGING_TIMESTAMP, T_OLD.to_string());
+        let existing = HashMap::from([
+            (AMZ_OBJECT_TAGGING.to_string(), "site=local".to_string()),
+            ("X-Minio-Internal-Tagging-Timestamp".to_string(), T_LOCAL.to_string()),
+        ]);
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(T_OLD)),
+            ..Default::default()
+        };
+
+        assert!(merge_replication_metadata_lww(&mut inbound, &existing, &opts));
+        assert_eq!(inbound.get(AMZ_OBJECT_TAGGING).map(String::as_str), Some("site=local"));
+        assert_eq!(
+            inbound.get("x-rustfs-internal-tagging-timestamp").map(String::as_str),
+            Some(T_LOCAL),
+            "the RustFS twin key must be materialized even when the source version only had the MinIO key"
+        );
+        assert_eq!(inbound.get("x-minio-internal-tagging-timestamp").map(String::as_str), Some(T_LOCAL));
+    }
+
+    /// When the inbound category wins, the stored timestamp must be the
+    /// source-authored one: the PUT path's eval_metadata stamps the
+    /// object-lock timestamps with the receiver's clock
+    /// (`parse_object_lock_retention`), which would otherwise make this
+    /// replica's clock the LWW authority and wedge later convergence.
+    #[tokio::test]
+    async fn inbound_win_pins_stored_timestamp_to_source_authored_value() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-retention-ts-pinned";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        // Local cleared retention at T_OLD.
+        let mut local = HashMap::new();
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_OLD.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        // Inbound newer retention: the source authored T_LOCAL, but the PUT
+        // path's eval_metadata stomped the metadata key with receiver-now
+        // (simulated by T_NEW here).
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "GOVERNANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2028-01-01T00:00:00Z".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_NEW.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_retention_timestamp: Some(parse_ts(T_LOCAL)),
+            ..versioned_opts(&version_id, inbound)
+        };
+        put_version(&set_disks, bucket, object, &version_id, &opts).await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP).as_deref(),
+            Some(T_LOCAL),
+            "the stored category timestamp must be the source-authored time, not the receiver's clock"
+        );
+        assert_eq!(info.user_defined.get(AMZ_OBJECT_LOCK_MODE_LOWER).map(String::as_str), Some("GOVERNANCE"));
+    }
+
+    #[tokio::test]
+    async fn newer_local_tag_deletion_survives_stale_inbound_tags() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-tagging-deleted";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        // Local DeleteObjectTagging state: no tags, but a newer tagging timestamp.
+        let mut local = HashMap::new();
+        insert_str(&mut local, SUFFIX_TAGGING_TIMESTAMP, T_LOCAL.to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_tagging_opts(&version_id, "site=remote", T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert!(
+            info.user_tags.is_empty(),
+            "a newer local tag deletion must not be resurrected by older inbound tags"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_TAGGING_TIMESTAMP).as_deref(), Some(T_LOCAL));
+    }
+
+    /// Destination version under an active legal hold at `hold_timestamp`,
+    /// plus an active COMPLIANCE retention (no retention timestamp).
+    async fn seed_locked_version(set_disks: &Arc<SetDisks>, bucket: &str, object: &str, version_id: &str, hold_timestamp: &str) {
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "ON".to_string());
+        insert_str(&mut local, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, hold_timestamp.to_string());
+        local.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "COMPLIANCE".to_string());
+        local.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2099-01-01T00:00:00Z".to_string());
+        put_version(set_disks, bucket, object, version_id, &versioned_opts(version_id, local)).await;
+    }
+
+    /// Inbound legal-hold release from a source that also carries the (same)
+    /// COMPLIANCE retention; the sender stamps a source timestamp for every
+    /// category the source version has.
+    fn inbound_legal_hold_release_opts(version_id: &str, timestamp: &str) -> ObjectOptions {
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "OFF".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, timestamp.to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "COMPLIANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2099-01-01T00:00:00Z".to_string());
+        insert_str(&mut inbound, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP, T_OLD.to_string());
+        ObjectOptions {
+            replication_request: true,
+            replication_legalhold_timestamp: Some(parse_ts(timestamp)),
+            replication_retention_timestamp: Some(parse_ts(T_OLD)),
+            ..versioned_opts(version_id, inbound)
+        }
+    }
+
+    /// The source's lock state governs the replica: a legal-hold release (or a
+    /// retention change) can only reach this site through the authorized
+    /// replication write, so the commit-time WORM gate must not reject it
+    /// because the destination version is currently locked.
+    #[tokio::test]
+    async fn inbound_newer_legal_hold_release_updates_locked_version() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-locked-release-newer";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_locked_version(&set_disks, bucket, object, &version_id, T_OLD).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_legal_hold_release_opts(&version_id, T_NEW),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str),
+            Some("OFF"),
+            "a newer source-side legal hold release must be applied to the locked replica"
+        );
+        assert_eq!(get_str(&info.user_defined, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP).as_deref(), Some(T_NEW));
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_MODE_LOWER).map(String::as_str),
+            Some("COMPLIANCE"),
+            "the untouched retention category must survive the write"
+        );
+    }
+
+    /// Skipping the WORM gate for replication writes must not weaken LWW: a
+    /// stale inbound release still loses to a hold applied more recently here.
+    #[tokio::test]
+    async fn inbound_stale_legal_hold_release_keeps_newer_local_hold() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-locked-release-stale";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_locked_version(&set_disks, bucket, object, &version_id, T_LOCAL).await;
+
+        put_version(
+            &set_disks,
+            bucket,
+            object,
+            &version_id,
+            &inbound_legal_hold_release_opts(&version_id, T_OLD),
+        )
+        .await;
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(
+            info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str),
+            Some("ON"),
+            "a stale inbound release must not lift a hold applied more recently on this site"
+        );
+        assert_eq!(
+            get_str(&info.user_defined, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP).as_deref(),
+            Some(T_LOCAL)
+        );
+    }
+
+    /// A replication write that carries no source decision for a locking
+    /// category (here: tags changed at a source that never held the object)
+    /// must not lift the destination's hold by replacing the metadata
+    /// unjudged; it stays WORM-rejected like a local overwrite.
+    #[tokio::test]
+    async fn inbound_without_legal_hold_timestamp_stays_rejected_on_held_version() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-locked-unjudged-category";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_locked_version(&set_disks, bucket, object, &version_id, T_OLD).await;
+
+        let mut inbound = HashMap::new();
+        inbound.insert(AMZ_OBJECT_TAGGING.to_string(), "k=v".to_string());
+        insert_str(&mut inbound, SUFFIX_TAGGING_TIMESTAMP, T_NEW.to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_MODE_LOWER.to_string(), "COMPLIANCE".to_string());
+        inbound.insert(AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE_LOWER.to_string(), "2099-01-01T00:00:00Z".to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_tagging_timestamp: Some(parse_ts(T_NEW)),
+            replication_retention_timestamp: Some(parse_ts(T_NEW)),
+            replication_legalhold_timestamp: None,
+            ..versioned_opts(&version_id, inbound)
+        };
+        let mut reader = PutObjReader::from_vec(b"lww-body".to_vec());
+        let err = set_disks
+            .put_object(bucket, object, &mut reader, &opts)
+            .await
+            .expect_err("a replication write without the legal-hold source timestamp must stay rejected");
+        assert!(matches!(err, StorageError::PrefixAccessDenied(_, _)), "unexpected error: {err}");
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str), Some("ON"));
+    }
+
+    /// The gate runs before the replication bypass, so malformed persisted
+    /// lock metadata still fails closed for an authorized replication write.
+    #[tokio::test]
+    async fn replication_write_on_malformed_lock_metadata_still_fails_closed() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-locked-malformed";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        let mut local = HashMap::new();
+        local.insert(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER.to_string(), "MAYBE".to_string());
+        put_version(&set_disks, bucket, object, &version_id, &versioned_opts(&version_id, local)).await;
+
+        let mut reader = PutObjReader::from_vec(b"lww-body".to_vec());
+        let err = set_disks
+            .put_object(bucket, object, &mut reader, &inbound_legal_hold_release_opts(&version_id, T_NEW))
+            .await
+            .expect_err("malformed persisted lock metadata must fail the replication write closed");
+        assert!(!matches!(err, StorageError::PrefixAccessDenied(_, _)), "unexpected error: {err}");
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str), Some("MAYBE"));
+    }
+
+    /// The bypass is scoped to authorized replication writes: the same
+    /// explicit-version PUT without `replication_request` stays WORM-rejected.
+    #[tokio::test]
+    async fn non_replication_overwrite_of_locked_version_is_still_rejected() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-locked-plain-put";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+        seed_locked_version(&set_disks, bucket, object, &version_id, T_OLD).await;
+
+        let opts = ObjectOptions {
+            replication_request: false,
+            ..inbound_legal_hold_release_opts(&version_id, T_NEW)
+        };
+        let mut reader = PutObjReader::from_vec(b"lww-body".to_vec());
+        let err = set_disks
+            .put_object(bucket, object, &mut reader, &opts)
+            .await
+            .expect_err("a non-replication overwrite of a locked version must stay rejected");
+        assert!(matches!(err, StorageError::PrefixAccessDenied(_, _)), "unexpected error: {err}");
+
+        let info = version_info(&set_disks, bucket, object, &version_id).await;
+        assert_eq!(info.user_defined.get(AMZ_OBJECT_LOCK_LEGAL_HOLD_LOWER).map(String::as_str), Some("ON"));
+    }
+}
+
 #[cfg(test)]
 mod inline_put_commit_path_tests {
     use super::hermetic_set_disks_support::hermetic_set_disks_isolated as hermetic_set_disks;
@@ -13873,6 +14535,72 @@ mod put_object_tmp_cleanup_tests {
         assert_eq!(body, original_body);
     }
 
+    /// A local PutObjectRetention / PutObjectLegalHold clear persists empty
+    /// lock keys (`parse_object_lock_retention`). The commit-time WORM gate
+    /// must read that as unlocked: an explicit-version PUT (the inbound
+    /// replication transport) and a version delete both have to succeed
+    /// (rustfs/backlog#1953).
+    #[tokio::test]
+    async fn explicit_version_overwrite_and_delete_succeed_after_local_lock_clear() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "put-explicit-version-cleared-lock";
+        let object = "object";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let mut initial_reader = PutObjReader::from_vec(b"original".to_vec());
+        let initial = set_disks
+            .put_object(
+                bucket,
+                object,
+                &mut initial_reader,
+                &ObjectOptions {
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("initial version should be written");
+        let version_id = initial
+            .version_id
+            .expect("versioned PUT should return a version ID")
+            .to_string();
+        let version_opts = ObjectOptions {
+            versioned: true,
+            version_id: Some(version_id.clone()),
+            delete_replication_config_snapshot: Some(Arc::new(DeleteReplicationConfigSnapshot::default())),
+            object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(ObjectLockConfigState::ConfirmedAbsent))),
+            ..Default::default()
+        };
+        set_disks
+            .put_object_metadata(
+                bucket,
+                object,
+                &ObjectOptions {
+                    eval_metadata: Some(HashMap::from([
+                        (X_AMZ_OBJECT_LOCK_MODE.as_str().to_string(), String::new()),
+                        (X_AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE.as_str().to_string(), String::new()),
+                        (X_AMZ_OBJECT_LOCK_LEGAL_HOLD.as_str().to_string(), String::new()),
+                    ])),
+                    ..version_opts.clone()
+                },
+            )
+            .await
+            .expect("cleared lock metadata should be written");
+
+        let mut replacement = PutObjReader::from_vec(b"replacement".to_vec());
+        set_disks
+            .put_object(bucket, object, &mut replacement, &version_opts)
+            .await
+            .expect("explicit-version PUT must not be wedged by cleared lock metadata");
+
+        set_disks
+            .delete_object(bucket, object, version_opts)
+            .await
+            .expect("version delete must not be wedged by cleared lock metadata");
+    }
+
     #[tokio::test]
     async fn version_only_copy_checks_the_destination_version_object_lock() {
         let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;

rustfs/src/app/object_usecase.rs

@@ -39,7 +39,7 @@ use super::storage_api::object_usecase::bucket::{
     metadata_sys,
     object_lock::{
         objectlock::{get_object_legalhold_meta, get_object_retention_meta},
-        objectlock_sys::{check_object_lock_for_deletion, is_retention_active},
+        objectlock_sys::{check_object_lock_for_deletion, is_retention_active, replication_write_may_pass_worm_gate},
     },
     predict_lifecycle_expiration,
     quota::{QuotaCheckResult, QuotaError, QuotaOperation},
@@ -3938,6 +3938,12 @@ pub(crate) fn validate_existing_object_lock_for_write(existing_obj_info: &Object
     if put_like_write_creates_new_version(opts) {
         return Ok(());
     }
+    // An authorized replication write may replace the locked version when the
+    // set layer's commit-lock LWW will judge every locking category; the set
+    // layer re-checks the same rule under the lock.
+    if replication_write_may_pass_worm_gate(&existing_obj_info.user_defined, opts) {
+        return Ok(());
+    }
 
     let legal_hold = get_object_legalhold_meta(&existing_obj_info.user_defined);
     if legal_hold
@@ -11037,6 +11043,47 @@ mod tests {
         assert_eq!(err.code(), &S3ErrorCode::AccessDenied);
     }
 
+    /// The source's lock state governs the replica (rustfs/backlog#1953):
+    /// an authorized replication write carrying the locking category's source
+    /// timestamp may overwrite a locked version; the set layer's LWW then
+    /// decides per category.
+    #[test]
+    fn validate_existing_object_lock_allows_authorized_replication_overwrite() {
+        let opts = ObjectOptions {
+            versioned: true,
+            version_id: Some(Uuid::new_v4().to_string()),
+            replication_request: true,
+            replication_retention_timestamp: Some(OffsetDateTime::UNIX_EPOCH),
+            replication_legalhold_timestamp: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        };
+
+        validate_existing_object_lock_for_write(&compliance_retained_object_info(), &opts)
+            .expect("replication write must bypass the destination COMPLIANCE lock");
+        validate_existing_object_lock_for_write(&legal_hold_object_info(), &opts)
+            .expect("replication write must bypass the destination legal hold");
+    }
+
+    /// Without the locking category's source timestamp the LWW merge cannot
+    /// judge it, so the write stays rejected instead of lifting the lock.
+    #[test]
+    fn validate_existing_object_lock_rejects_replication_overwrite_without_lock_timestamp() {
+        let opts = ObjectOptions {
+            versioned: true,
+            version_id: Some(Uuid::new_v4().to_string()),
+            replication_request: true,
+            replication_tagging_timestamp: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        };
+
+        let err = validate_existing_object_lock_for_write(&compliance_retained_object_info(), &opts)
+            .expect_err("COMPLIANCE lock must hold without a retention source timestamp");
+        assert_eq!(err.code(), &S3ErrorCode::AccessDenied);
+        let err = validate_existing_object_lock_for_write(&legal_hold_object_info(), &opts)
+            .expect_err("legal hold must hold without a legal-hold source timestamp");
+        assert_eq!(err.code(), &S3ErrorCode::AccessDenied);
+    }
+
     #[test]
     fn is_put_object_extract_requested_accepts_meta_header() {
         let mut headers = HeaderMap::new();

rustfs/src/app/storage_api.rs

@@ -587,6 +587,16 @@ pub(crate) mod bucket {
                     retain_until_date,
                 )
             }
+
+            pub(crate) fn replication_write_may_pass_worm_gate(
+                user_defined: &std::collections::HashMap<String, String>,
+                opts: &crate::storage::storage_api::StorageObjectOptions,
+            ) -> bool {
+                crate::storage::storage_api::ecstore_bucket::object_lock::objectlock_sys::replication_write_may_pass_worm_gate(
+                    user_defined,
+                    opts,
+                )
+            }
         }
     }
 

rustfs/src/storage/options.rs

@@ -554,10 +554,11 @@ fn apply_replication_timestamps_from_headers(headers: &HeaderMap<HeaderValue>, o
 
     // 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.
+    // times instead of falling back to mod_time. Receiver-side LWW happens at
+    // the set layer under the object write lock
+    // (ecstore set_disk::ops::object::merge_replication_metadata_lww,
+    // rustfs/backlog#1953): a category whose stored timestamp is newer than
+    // the inbound one keeps the local values.
     for (timestamp, suffix) in [
         (opts.replication_tagging_timestamp, SUFFIX_TAGGING_TIMESTAMP),
         (opts.replication_retention_timestamp, SUFFIX_OBJECTLOCK_RETENTION_TIMESTAMP),