fix(replication): apply receiver-side LWW to inbound metadata categories (#6379)

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_extension_test.rs

@@ -57,7 +57,7 @@ use rustfs_madmin::{
     AddServiceAccountReq, ListServiceAccountsResp, PeerInfo, PeerSite, ReplicateAddStatus, ReplicateEditStatus,
     ReplicateRemoveStatus, SRRemoveReq, SRResyncOpStatus, SRStatusInfo, SiteReplicationInfo, SyncStatus,
 };
-use s3s::header::X_AMZ_REPLICATION_STATUS;
+use s3s::header::{X_AMZ_REPLICATION_STATUS, X_AMZ_TAGGING};
 use sha2::{Digest, Sha256};
 use std::collections::BTreeMap;
 use std::convert::Infallible;
@@ -2023,6 +2023,7 @@ async fn forward_replication_proxy_request(
     client: &reqwest::Client,
     request_count: &AtomicU64,
     mut replication_enabled: watch::Receiver<bool>,
+    mut held_tagging: watch::Receiver<Option<String>>,
 ) -> Response<Full<bytes::Bytes>> {
     let (parts, body) = request.into_parts();
     let is_replication = parts
@@ -2036,6 +2037,17 @@ async fn forward_replication_proxy_request(
                 return proxy_error_response("replication gate closed");
             }
         }
+        // Content-keyed hold: park only the replication request whose
+        // `x-amz-tagging` matches the held value, letting every other delivery
+        // through, so a test can make one specific (stale) delivery the last
+        // write the backend sees.
+        if let Some(tagging) = parts.headers.get(X_AMZ_TAGGING).and_then(|value| value.to_str().ok()) {
+            while held_tagging.borrow().as_deref() == Some(tagging) {
+                if held_tagging.changed().await.is_err() {
+                    return proxy_error_response("replication tag hold closed");
+                }
+            }
+        }
     }
 
     let Some(path_and_query) = parts.uri.path_and_query() else {
@@ -2070,12 +2082,26 @@ async fn start_replication_counting_proxy(
     backend_url: &str,
     tasks: &mut JoinSet<()>,
 ) -> Result<(String, Arc<AtomicU64>, watch::Sender<bool>), Box<dyn Error + Send + Sync>> {
+    let (proxy_url, request_count, replication_enabled, _held_tagging) =
+        start_replication_counting_proxy_with_tag_hold(backend_url, tasks).await?;
+    Ok((proxy_url, request_count, replication_enabled))
+}
+
+/// [`start_replication_counting_proxy`] plus a content-keyed hold: while the
+/// returned `watch::Sender<Option<String>>` holds `Some(tagging)`, replication
+/// requests whose `x-amz-tagging` equals `tagging` are parked (and still
+/// counted); all other traffic flows. Send `None` to release them.
+async fn start_replication_counting_proxy_with_tag_hold(
+    backend_url: &str,
+    tasks: &mut JoinSet<()>,
+) -> Result<(String, Arc<AtomicU64>, watch::Sender<bool>, watch::Sender<Option<String>>), Box<dyn Error + Send + Sync>> {
     let listener = TcpListener::bind("127.0.0.1:0").await?;
     let proxy_url = format!("http://{}", listener.local_addr()?);
     let backend_url = backend_url.to_string();
     let request_count = Arc::new(AtomicU64::new(0));
     let task_request_count = request_count.clone();
     let (replication_enabled, task_replication_enabled) = watch::channel(true);
+    let (held_tagging, task_held_tagging) = watch::channel(None);
     tasks.spawn(async move {
         let client = local_http_client();
         let mut connections = JoinSet::new();
@@ -2087,12 +2113,14 @@ async fn start_replication_counting_proxy(
                     let client = client.clone();
                     let request_count = task_request_count.clone();
                     let replication_enabled = task_replication_enabled.clone();
+                    let held_tagging = task_held_tagging.clone();
                     connections.spawn(async move {
                         let service = service_fn(move |request| {
                             let backend_url = backend_url.clone();
                             let client = client.clone();
                             let request_count = request_count.clone();
                             let replication_enabled = replication_enabled.clone();
+                            let held_tagging = held_tagging.clone();
                             async move {
                                 Ok::<_, Infallible>(
                                     forward_replication_proxy_request(
@@ -2101,6 +2129,7 @@ async fn start_replication_counting_proxy(
                                         &client,
                                         &request_count,
                                         replication_enabled,
+                                        held_tagging,
                                     )
                                     .await,
                                 )
@@ -2113,7 +2142,7 @@ async fn start_replication_counting_proxy(
             }
         }
     });
-    Ok((proxy_url, request_count, replication_enabled))
+    Ok((proxy_url, request_count, replication_enabled, held_tagging))
 }
 
 async fn site_replication_remove(
@@ -6949,6 +6978,395 @@ async fn test_site_replication_active_active_converges_without_loops_real_dual_n
     }
 }
 
+/// Replication status a site reports for one object version via HEAD
+/// (`x-amz-replication-status`), or `None` when the header is absent.
+async fn head_replication_status(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+) -> Result<Option<String>, Box<dyn Error + Send + Sync>> {
+    let head = client
+        .head_object()
+        .bucket(bucket)
+        .key(key)
+        .version_id(version_id)
+        .send()
+        .await?;
+    Ok(head.replication_status().map(|status| status.as_str().to_string()))
+}
+
+/// Poll one site until the version's replication status is one of `expected`.
+async fn wait_for_version_replication_status(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    expected: &[&str],
+    site: &str,
+) -> Result<String, Box<dyn Error + Send + Sync>> {
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(60);
+    loop {
+        let last = head_replication_status(client, bucket, key, version_id).await?;
+        if let Some(status) = last.as_deref()
+            && expected.contains(&status)
+        {
+            return Ok(status.to_string());
+        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!(
+                "{site}: {bucket}/{key}?versionId={version_id} replication status {last:?} never reached {expected:?}"
+            )
+            .into());
+        }
+        sleep(Duration::from_millis(200)).await;
+    }
+}
+
+async fn put_single_tag(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    tag_key: &str,
+    tag_value: &str,
+) -> Result<(), Box<dyn Error + Send + Sync>> {
+    client
+        .put_object_tagging()
+        .bucket(bucket)
+        .key(key)
+        .version_id(version_id)
+        .tagging(
+            aws_sdk_s3::types::Tagging::builder()
+                .tag_set(aws_sdk_s3::types::Tag::builder().key(tag_key).value(tag_value).build()?)
+                .build()?,
+        )
+        .send()
+        .await?;
+    Ok(())
+}
+
+async fn get_single_tag(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    tag_key: &str,
+) -> Result<Option<String>, Box<dyn Error + Send + Sync>> {
+    let tagging = client
+        .get_object_tagging()
+        .bucket(bucket)
+        .key(key)
+        .version_id(version_id)
+        .send()
+        .await?;
+    Ok(tagging
+        .tag_set()
+        .iter()
+        .find(|tag| tag.key() == tag_key)
+        .map(|tag| tag.value().to_string()))
+}
+
+/// Poll one site until the version's `tag_key` equals `expected`.
+async fn wait_for_single_tag(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    tag_key: &str,
+    expected: &str,
+    site: &str,
+) -> Result<(), Box<dyn Error + Send + Sync>> {
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(60);
+    loop {
+        let observed = get_single_tag(client, bucket, key, version_id, tag_key).await?;
+        if observed.as_deref() == Some(expected) {
+            return Ok(());
+        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!(
+                "{site}: {bucket}/{key}?versionId={version_id} tag {tag_key}={observed:?} never became {expected}"
+            )
+            .into());
+        }
+        sleep(Duration::from_millis(200)).await;
+    }
+}
+
+/// Tag key the dual-node LWW scenario edits on both sites.
+const LWW_TAG_KEY: &str = "owner";
+
+/// Assert the version's [`LWW_TAG_KEY`] stays `expected` on both sites for a
+/// full quiet window (no late stale delivery flips it back).
+async fn assert_tag_stable_on_both_sites(
+    site_a_client: &Client,
+    site_b_client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    expected: &str,
+    quiet: Duration,
+) -> Result<(), Box<dyn Error + Send + Sync>> {
+    let deadline = tokio::time::Instant::now() + quiet;
+    loop {
+        let on_a = get_single_tag(site_a_client, bucket, key, version_id, LWW_TAG_KEY).await?;
+        let on_b = get_single_tag(site_b_client, bucket, key, version_id, LWW_TAG_KEY).await?;
+        assert_eq!(on_a.as_deref(), Some(expected), "site A tag {LWW_TAG_KEY} regressed from the LWW winner");
+        assert_eq!(on_b.as_deref(), Some(expected), "site B tag {LWW_TAG_KEY} regressed from the LWW winner");
+        if tokio::time::Instant::now() >= deadline {
+            return Ok(());
+        }
+        sleep(Duration::from_millis(250)).await;
+    }
+}
+
+/// Wait until the counting proxy in front of a site has admitted `expected`
+/// replication requests in total (requests held by a closed gate still count).
+async fn wait_for_proxy_replication_requests(
+    counter: &AtomicU64,
+    expected: u64,
+    site: &str,
+) -> Result<(), Box<dyn Error + Send + Sync>> {
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(60);
+    loop {
+        let observed = counter.load(Ordering::Relaxed);
+        if observed >= expected {
+            return Ok(());
+        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!("{site} proxy saw {observed} replication requests, expected at least {expected}").into());
+        }
+        sleep(Duration::from_millis(25)).await;
+    }
+}
+
+/// rustfs/backlog#1953 (audit A4/P1-6): receiver-side LWW for replicated
+/// metadata categories, exercised end to end over the real dual-node
+/// active-active site-replication control plane — sender, worker, status
+/// bookkeeping and persisted failure recovery all participate (the single-server
+/// `replication_lww_receiver_test` only injects authorized replication PUTs).
+///
+/// Scenario on one versioned object:
+/// 1. reciprocal tag edits in real order (A then B) converge both sites on the
+///    newer tag and leave the author COMPLETED / the receiver REPLICA;
+/// 2. out-of-order delivery: A's edit is held at B's inbound proxy while B
+///    authors a newer edit that reaches A first; releasing the stale delivery
+///    must NOT roll B back — both sites settle on B's value and stay there
+///    through a quiet window, with no FAILED/PENDING status left behind;
+/// 3. persisted retry: B is stopped, A's delivery reaches FAILED, A restarts,
+///    then B returns and the scanner-replayed edit converges both sites forward.
+/// Durable metadata-MRF serialization/reconstruction is covered separately by
+/// `metadata_mrf_roundtrip_preserves_tags_and_admitted_targets`.
+#[tokio::test]
+async fn test_site_replication_tagging_lww_converges_active_active_real_dual_node() -> TestResult {
+    init_logging();
+
+    match tokio::time::timeout(Duration::from_secs(420), async {
+        // The scanner is fast for the final persisted-failure recovery phase.
+        // Step 2 finishes and proves a quiet stable winner before that phase,
+        // so a later scanner pass cannot mask its stale-delivery assertion.
+        let mut site_env = replication_fast_env();
+        site_env.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+        site_env.extend_from_slice(FAST_SCANNER_ENV);
+
+        let mut site_a_env = RustFSTestEnvironment::new().await?;
+        site_a_env.start_rustfs_server_with_env(vec![], &site_env).await?;
+
+        let mut site_b_env = RustFSTestEnvironment::new().await?;
+        site_b_env.start_rustfs_server_with_env(vec![], &site_env).await?;
+
+        let mut proxy_tasks = JoinSet::new();
+        let (site_a_proxy, site_a_replication_requests, _site_a_replication_enabled, site_a_held_tagging) =
+            start_replication_counting_proxy_with_tag_hold(&site_a_env.url, &mut proxy_tasks).await?;
+        let (site_b_proxy, site_b_replication_requests, _site_b_replication_enabled, site_b_held_tagging) =
+            start_replication_counting_proxy_with_tag_hold(&site_b_env.url, &mut proxy_tasks).await?;
+
+        let site_a_client = site_a_env.create_s3_client();
+        let site_b_client = site_b_env.create_s3_client();
+        let bucket = "site-repl-tag-lww";
+        let key = "lww.txt";
+
+        let add_status = site_replication_add(
+            &site_a_env,
+            &[
+                PeerSite {
+                    name: "lww-site-a".to_string(),
+                    endpoint: site_a_env.url.clone(),
+                    access_key: site_a_env.access_key.clone(),
+                    secret_key: site_a_env.secret_key.clone(),
+                    ..Default::default()
+                },
+                PeerSite {
+                    name: "lww-site-b".to_string(),
+                    endpoint: site_b_env.url.clone(),
+                    access_key: site_b_env.access_key.clone(),
+                    secret_key: site_b_env.secret_key.clone(),
+                    ..Default::default()
+                },
+            ],
+        )
+        .await?;
+        assert!(add_status.success, "unexpected site add result: {add_status:?}");
+
+        let site_info = wait_for_site_replication_enabled(&site_a_env, 2).await?;
+        wait_for_site_replication_enabled(&site_b_env, 2).await?;
+
+        // Route both directions through the counting proxies so inbound
+        // replication to B can be held (out-of-order delivery) and observed.
+        for (env_url, proxy_url, label) in [(&site_a_env.url, &site_a_proxy, "A"), (&site_b_env.url, &site_b_proxy, "B")] {
+            let mut peer = site_info
+                .sites
+                .iter()
+                .find(|peer| peer.endpoint == *env_url)
+                .ok_or_else(|| format!("site {label} peer missing from replication info"))?
+                .clone();
+            peer.endpoint = proxy_url.clone();
+            peer.sync_state = SyncStatus::Enable;
+            let edit = site_replication_edit(&site_a_env, "", &peer).await?;
+            assert!(edit.success, "unexpected site {label} endpoint edit: {edit:?}");
+        }
+        for env in [&site_a_env, &site_b_env] {
+            wait_for_site_replication_info(env, |info| {
+                info.sites.iter().any(|peer| peer.endpoint == site_a_proxy)
+                    && info.sites.iter().any(|peer| peer.endpoint == site_b_proxy)
+            })
+            .await?;
+        }
+
+        site_a_client.create_bucket().bucket(bucket).send().await?;
+        wait_for_bucket_on_target(&site_b_client, bucket).await?;
+
+        let version_id = site_a_client
+            .put_object()
+            .bucket(bucket)
+            .key(key)
+            .body(ByteStream::from_static(b"tag lww payload"))
+            .send()
+            .await?
+            .version_id()
+            .ok_or("site A PUT omitted version ID")?
+            .to_string();
+        wait_for_replicated_object(&site_b_client, bucket, key, "tag lww payload").await?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["COMPLETED"], "site A").await?;
+        wait_for_version_replication_status(&site_b_client, bucket, key, &version_id, &["REPLICA"], "site B").await?;
+        wait_for_proxy_replication_requests(&site_b_replication_requests, 1, "site B").await?;
+
+        // --- 1. reciprocal edits in real order: A then B ----------------------
+        put_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY, "a1").await?;
+        wait_for_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY, "a1", "site B").await?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["COMPLETED"], "site A").await?;
+        wait_for_version_replication_status(&site_b_client, bucket, key, &version_id, &["REPLICA"], "site B").await?;
+        wait_for_proxy_replication_requests(&site_b_replication_requests, 2, "site B").await?;
+
+        put_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY, "b1").await?;
+        wait_for_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY, "b1", "site A").await?;
+        wait_for_version_replication_status(&site_b_client, bucket, key, &version_id, &["COMPLETED"], "site B").await?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["REPLICA"], "site A").await?;
+        assert_tag_stable_on_both_sites(&site_a_client, &site_b_client, bucket, key, &version_id, "b1", Duration::from_secs(3))
+            .await?;
+
+        // --- 2. concurrent edits, stale delivery last ------------------------
+        // Both sites edit the same version while each other's delivery is
+        // parked at the peer's inbound proxy (content-keyed: only the
+        // `owner=a2` / `owner=b2` replication PUTs wait, everything else
+        // flows). B's edit is the newer one. Releasing A's stale `a2` first
+        // makes it the last write B sees while A itself still holds `a2`, so
+        // nothing A could re-deliver carries the winner: only receiver-side
+        // LWW on B can keep `b2`. Releasing `b2` afterwards converges A.
+        site_b_held_tagging.send(Some("owner=a2".to_string()))?;
+        site_a_held_tagging.send(Some("owner=b2".to_string()))?;
+        let a2_parked_at = site_b_replication_requests.load(Ordering::Relaxed) + 1;
+        let b2_parked_at = site_a_replication_requests.load(Ordering::Relaxed) + 1;
+        put_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY, "a2").await?;
+        wait_for_proxy_replication_requests(&site_b_replication_requests, a2_parked_at, "site B").await?;
+        sleep(Duration::from_millis(50)).await;
+        put_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY, "b2").await?;
+        wait_for_proxy_replication_requests(&site_a_replication_requests, b2_parked_at, "site A").await?;
+        assert_eq!(
+            get_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY)
+                .await?
+                .as_deref(),
+            Some("a2")
+        );
+        assert_eq!(
+            get_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY)
+                .await?
+                .as_deref(),
+            Some("b2")
+        );
+
+        // Release the stale a2 delivery onto B: the newer local b2 must
+        // survive, and the delivery itself must still succeed (A reaches
+        // COMPLETED instead of looping through MRF with the stale value).
+        site_b_held_tagging.send(None)?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["COMPLETED"], "site A").await?;
+        let stale_deadline = tokio::time::Instant::now() + Duration::from_secs(3);
+        loop {
+            assert_eq!(
+                get_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY)
+                    .await?
+                    .as_deref(),
+                Some("b2"),
+                "a stale inbound delivery rolled back site B's newer tag (receiver-side LWW regression)"
+            );
+            if tokio::time::Instant::now() >= stale_deadline {
+                break;
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+
+        // Release b2 onto A: the newer edit wins there and both sites settle.
+        // B's own version may legitimately read REPLICA here: the stale inbound
+        // a2 write re-labelled it as a replica write (keeping B's tags); what
+        // must not remain is PENDING/FAILED.
+        site_a_held_tagging.send(None)?;
+        wait_for_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY, "b2", "site A").await?;
+        wait_for_version_replication_status(&site_b_client, bucket, key, &version_id, &["COMPLETED", "REPLICA"], "site B")
+            .await?;
+        assert_tag_stable_on_both_sites(&site_a_client, &site_b_client, bucket, key, &version_id, "b2", Duration::from_secs(4))
+            .await?;
+        for (client, site) in [(&site_a_client, "site A"), (&site_b_client, "site B")] {
+            let status = head_replication_status(client, bucket, key, &version_id).await?;
+            assert!(
+                matches!(status.as_deref(), Some("COMPLETED" | "REPLICA")),
+                "{site} must not be left PENDING/FAILED after the concurrent edits: {status:?}"
+            );
+        }
+
+        // --- 3. persisted FAILED state survives a source restart ------------
+        site_b_env.stop_server();
+        put_single_tag(&site_a_client, bucket, key, &version_id, LWW_TAG_KEY, "a3").await?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["FAILED"], "site A").await?;
+        site_a_env.restart_server_preserving_data(vec![], &site_env).await?;
+        wait_for_site_replication_enabled(&site_a_env, 2).await?;
+        site_b_env.restart_server_preserving_data(vec![], &site_env).await?;
+        wait_for_site_replication_enabled(&site_b_env, 2).await?;
+
+        wait_for_single_tag(&site_b_client, bucket, key, &version_id, LWW_TAG_KEY, "a3", "site B").await?;
+        wait_for_version_replication_status(&site_a_client, bucket, key, &version_id, &["COMPLETED"], "site A").await?;
+        wait_for_version_replication_status(&site_b_client, bucket, key, &version_id, &["REPLICA"], "site B").await?;
+        assert_tag_stable_on_both_sites(&site_a_client, &site_b_client, bucket, key, &version_id, "a3", Duration::from_secs(3))
+            .await?;
+
+        // The object itself never forked: one version on each side.
+        tokio::time::timeout(
+            Duration::from_secs(70),
+            assert_replication_converged(&site_a_client, bucket, &site_b_client, bucket),
+        )
+        .await??;
+        let state = list_replication_state(&site_a_client, bucket).await?;
+        assert_eq!(state.len(), 1, "tag edits must not create new object versions: {state:?}");
+        assert_eq!(state[0].version_id, version_id);
+
+        proxy_tasks.abort_all();
+        Ok(())
+    })
+    .await
+    {
+        Ok(result) => result,
+        Err(_) => Err("site replication tagging LWW test timed out".into()),
+    }
+}
 #[tokio::test]
 async fn test_site_replication_replicates_policy_backed_user_access_real_dual_node() -> Result<(), Box<dyn Error + Send + Sync>> {
     init_logging();

crates/e2e_test/src/replication_lww_receiver_test.rs

@@ -0,0 +1,155 @@
+#![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).
+//!
+//! The real dual-site path (sender, worker, status bookkeeping, MRF replay)
+//! is covered by
+//! `replication_extension_test::test_site_replication_tagging_lww_converges_active_active_real_dual_node`;
+//! this file stays as the fast, single-process receiver check.
+
+use crate::common::{RustFSTestEnvironment, init_logging};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+
+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")]
+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/bucket/bucket_target_sys.rs

@@ -58,8 +58,8 @@ use rustfs_config::{DEFAULT_TRUST_LEAF_CERT_AS_CA, ENV_TRUST_LEAF_CERT_AS_CA, RU
 use rustfs_utils::egress::{OutboundUrlError, validate_outbound_url};
 use rustfs_utils::http::{
     AMZ_BUCKET_REPLICATION_STATUS, AMZ_OBJECT_LOCK_BYPASS_GOVERNANCE, AMZ_OBJECT_LOCK_LEGAL_HOLD, AMZ_OBJECT_LOCK_MODE,
-    AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE, AMZ_STORAGE_CLASS, AMZ_WEBSITE_REDIRECT_LOCATION, is_amz_header, is_minio_header,
-    is_rustfs_header, is_standard_header, is_storageclass_header,
+    AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE, AMZ_OBJECT_TAGGING_LOWER, AMZ_STORAGE_CLASS, AMZ_WEBSITE_REDIRECT_LOCATION, is_amz_header,
+    is_minio_header, is_rustfs_header, is_standard_header, is_storageclass_header,
 };
 use rustfs_utils::http::{
     SUFFIX_FORCE_DELETE, SUFFIX_SOURCE_DELETEMARKER, SUFFIX_SOURCE_ETAG, SUFFIX_SOURCE_MTIME, SUFFIX_SOURCE_PROXY_REQUEST,
@@ -1774,6 +1774,22 @@ impl PutObjectOptions {
             Self::insert_checked(&mut header, AMZ_BUCKET_REPLICATION_STATUS, self.internal.replication_status.as_str());
         }
 
+        // MinIO PutObjectOptions.Header parity: object tags travel on the
+        // `x-amz-tagging` header (form-urlencoded). `replication_put_object_options`
+        // fills `user_tags` from the source version; without this header the
+        // whole-object transport delivered a tagless replica, so tag edits
+        // never reached the peer and the receiver-side LWW comparison
+        // (rustfs/backlog#1953) had nothing to judge.
+        if !self.user_tags.is_empty() {
+            let mut tags: Vec<(&String, &String)> = self.user_tags.iter().collect();
+            tags.sort();
+            let mut encoded = url::form_urlencoded::Serializer::new(String::new());
+            for (key, value) in tags {
+                encoded.append_pair(key, value);
+            }
+            Self::insert_checked(&mut header, AMZ_OBJECT_TAGGING_LOWER, &encoded.finish());
+        }
+
         for (k, v) in &self.user_metadata {
             let Ok(header_value) = HeaderValue::from_str(v) else {
                 warn!("skipping user metadata header with invalid value: {}", k);
@@ -3195,6 +3211,29 @@ mod tests {
         }
     }
 
+    #[test]
+    fn put_object_headers_carry_user_tags_on_x_amz_tagging() {
+        // rustfs/backlog#1953: tag edits replicate through the whole-object
+        // transport, so the source tags must travel on x-amz-tagging.
+        let mut opts = PutObjectOptions::default();
+        opts.user_tags.insert("owner".to_string(), "site a".to_string());
+        opts.user_tags.insert("env".to_string(), "prod".to_string());
+
+        let header = opts.header();
+        let tagging = header
+            .get(AMZ_OBJECT_TAGGING_LOWER)
+            .expect("user tags must be transported on x-amz-tagging")
+            .to_str()
+            .expect("tag header must be ASCII");
+        // Deterministic key order; values are form-urlencoded.
+        assert_eq!(tagging, "env=prod&owner=site+a");
+
+        assert!(
+            PutObjectOptions::default().header().get(AMZ_OBJECT_TAGGING_LOWER).is_none(),
+            "a tagless source must not send an empty x-amz-tagging header"
+        );
+    }
+
     #[test]
     fn put_object_headers_omit_unset_replication_timestamps() {
         // UNIX_EPOCH means "never modified on the source"; sending it would

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

@@ -4194,6 +4194,30 @@ mod tests {
         assert_eq!(ri.checksum, Some(checksum));
     }
 
+    #[test]
+    fn metadata_mrf_roundtrip_preserves_tags_and_admitted_targets() {
+        let target = "arn:rustfs:replication:target-a";
+        let object = ObjectInfo {
+            bucket: "source".to_string(),
+            name: "object".to_string(),
+            version_id: Some(Uuid::new_v4()),
+            user_tags: Arc::new("owner=a3".to_string()),
+            ..Default::default()
+        };
+        let live =
+            replicate_object_info_from_object_info(object.clone(), test_replicate_decision(&[target]), ReplicationType::Metadata);
+        let persisted = live.to_mrf_entry();
+        let encoded = encode_mrf_file(std::slice::from_ref(&persisted)).expect("metadata MRF entry should encode");
+        let decoded = decode_mrf_file(&encoded).expect("metadata MRF entry should decode");
+
+        assert_eq!(decoded[0].op, MrfOpKind::Metadata);
+        assert_eq!(decoded[0].target_arns, vec![target.to_string()]);
+        let replayed = admitted_mrf_replicate_object(object, &decoded[0], ReplicationType::Metadata);
+        assert_eq!(replayed.op_type, ReplicationType::Metadata);
+        assert_eq!(replayed.user_tags, "owner=a3");
+        assert_eq!(replayed.admitted_target_arns(), vec![target.to_string()]);
+    }
+
     #[tokio::test]
     async fn mrf_save_admission_waits_for_capacity_instead_of_dropping() {
         let (tx, mut rx) = mpsc::channel(1);

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

@@ -76,7 +76,8 @@ use metrics::counter;
 use rmp_serde;
 use rustfs_s3_types::EventName;
 use rustfs_utils::http::{
-    AMZ_TAGGING_DIRECTIVE, SUFFIX_REPLICATION_RESET, SUFFIX_REPLICATION_STATUS, has_internal_suffix, insert_str,
+    AMZ_BUCKET_REPLICATION_STATUS, AMZ_TAGGING_DIRECTIVE, SUFFIX_REPLICATION_RESET, SUFFIX_REPLICATION_STATUS,
+    has_internal_suffix, insert_str,
 };
 use rustfs_utils::{DEFAULT_SIP_HASH_KEY, get_env_usize, sip_hash};
 #[cfg(test)]
@@ -174,6 +175,14 @@ fn has_raw_status(err: &SdkError<HeadObjectError>, status: u16) -> bool {
     err.raw_response().is_some_and(|r| r.status().as_u16() == status)
 }
 
+fn metadata_requires_existing_target(op_type: ReplicationType, object_info: &ObjectInfo) -> bool {
+    op_type == ReplicationType::Metadata
+        && object_info
+            .user_defined
+            .get(AMZ_BUCKET_REPLICATION_STATUS)
+            .is_some_and(|status| status.eq_ignore_ascii_case(ReplicationStatusType::Replica.as_str()))
+}
+
 const METRIC_VERSION_IDENTITY_DRIFT_TOTAL: &str = "rustfs_replication_version_identity_drift_total";
 
 /// Targets that already produced a version-identity-drift warning this
@@ -3494,6 +3503,7 @@ async fn resolve_replicate_all_action(
         start_time,
         ssec_audit_required,
     } = ctx;
+    let require_existing_target = metadata_requires_existing_target(roi.op_type, &object_info);
     let replication_action;
     match head_object_for_worker(tgt_client.as_ref(), &tgt_client.bucket, object, roi.version_id.map(|v| v.to_string())).await {
         Ok(oi) => {
@@ -3555,7 +3565,13 @@ async fn resolve_replicate_all_action(
                 // Version-ID format mismatch: retry without versionId and compare ETags.
                 match head_object_fallback(tgt_client, object).await {
                     Ok(Some(oi)) => {
-                        replication_action = if replication_etags_match(object_info.etag.as_deref(), oi.e_tag.as_deref()) {
+                        let etags_match = replication_etags_match(object_info.etag.as_deref(), oi.e_tag.as_deref());
+                        if require_existing_target && !etags_match {
+                            rinfo.error = Some("replica metadata target does not contain matching object data".to_string());
+                            rinfo.duration = (OffsetDateTime::now_utc() - start_time).unsigned_abs();
+                            return None;
+                        }
+                        replication_action = if etags_match {
                             if ssec_audit_required
                                 && !settle_ssec_passthrough_evidence(&oi, tgt_client, bucket, object, rinfo).await
                             {
@@ -3568,6 +3584,11 @@ async fn resolve_replicate_all_action(
                         };
                     }
                     Ok(None) => {
+                        if require_existing_target {
+                            rinfo.error = Some("replica metadata target does not contain this object version".to_string());
+                            rinfo.duration = (OffsetDateTime::now_utc() - start_time).unsigned_abs();
+                            return None;
+                        }
                         replication_action = ReplicationAction::All;
                     }
                     Err(e2) => {
@@ -3593,7 +3614,12 @@ async fn resolve_replicate_all_action(
                         return None;
                     }
                 }
-            } else if e.as_service_error().is_some_and(|se| se.is_not_found()) {
+            } else if e.as_service_error().is_some_and(|se| se.is_not_found()) || has_raw_status(&e, 404) {
+                if require_existing_target {
+                    rinfo.error = Some("replica metadata target does not contain this object version".to_string());
+                    rinfo.duration = (OffsetDateTime::now_utc() - start_time).unsigned_abs();
+                    return None;
+                }
                 replication_action = ReplicationAction::All;
             } else {
                 rinfo.error = Some(e.to_string());
@@ -3868,6 +3894,7 @@ 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
@@ -3921,6 +3948,113 @@ mod tests {
         })
     }
 
+    fn spawn_head_status_server(status: u16) -> (String, std::thread::JoinHandle<()>) {
+        use std::io::{Read, Write};
+
+        let listener = std::net::TcpListener::bind(("127.0.0.1", 0)).expect("test HTTP listener should bind");
+        let endpoint = format!("http://{}", listener.local_addr().expect("test HTTP listener should have an address"));
+        let handle = std::thread::spawn(move || {
+            let (mut stream, _) = listener.accept().expect("test HTTP client should connect");
+            let mut request = [0_u8; 8192];
+            let bytes_read = stream.read(&mut request).expect("test HTTP request should be read");
+            assert!(bytes_read > 0, "test HTTP request should not be empty");
+            assert!(request[..bytes_read].starts_with(b"HEAD "), "replication comparison must use HEAD");
+            write!(stream, "HTTP/1.1 {status} Test\r\nContent-Length: 0\r\nConnection: close\r\n\r\n")
+                .expect("test HTTP response should be written");
+        });
+        (endpoint, handle)
+    }
+
+    #[tokio::test]
+    async fn replica_metadata_missing_target_stops_before_full_put() {
+        let (endpoint, server) = spawn_head_status_server(404);
+        let target = test_target_client(endpoint);
+        let roi = ReplicateObjectInfo {
+            bucket: "source".to_string(),
+            name: "object".to_string(),
+            version_id: Some(Uuid::new_v4()),
+            op_type: ReplicationType::Metadata,
+            // Normal metadata writes replace REPLICA with per-target PENDING
+            // before constructing the worker request.
+            replication_status: ReplicationStatusType::Pending,
+            ..Default::default()
+        };
+        let object_info = ObjectInfo {
+            bucket: roi.bucket.clone(),
+            name: roi.name.clone(),
+            version_id: roi.version_id,
+            etag: Some("source-etag".to_string()),
+            user_defined: Arc::new(HashMap::from([(
+                AMZ_BUCKET_REPLICATION_STATUS.to_string(),
+                ReplicationStatusType::Replica.as_str().to_string(),
+            )])),
+            ..Default::default()
+        };
+        let mut rinfo = replicate_all_target_info(&roi, &target);
+
+        let action = resolve_replicate_all_action(
+            ReplicateAllActionContext {
+                roi: &roi,
+                tgt_client: &target,
+                bucket: &roi.bucket,
+                object: &roi.name,
+                start_time: OffsetDateTime::now_utc(),
+                ssec_audit_required: false,
+            },
+            object_info,
+            &mut rinfo,
+        )
+        .await;
+
+        assert!(action.is_none(), "missing replica metadata targets must not reach the payload PUT path");
+        assert_eq!(rinfo.replication_status, ReplicationStatusType::Failed);
+        assert_eq!(
+            rinfo.error.as_deref(),
+            Some("replica metadata target does not contain this object version")
+        );
+        server.join().expect("test HTTP server should finish");
+    }
+
+    #[tokio::test]
+    async fn source_metadata_missing_target_rebuilds_object() {
+        let (endpoint, server) = spawn_head_status_server(404);
+        let target = test_target_client(endpoint);
+        let roi = ReplicateObjectInfo {
+            bucket: "source".to_string(),
+            name: "object".to_string(),
+            version_id: Some(Uuid::new_v4()),
+            op_type: ReplicationType::Metadata,
+            replication_status: ReplicationStatusType::Pending,
+            ..Default::default()
+        };
+        let object_info = ObjectInfo {
+            bucket: roi.bucket.clone(),
+            name: roi.name.clone(),
+            version_id: roi.version_id,
+            etag: Some("source-etag".to_string()),
+            ..Default::default()
+        };
+        let mut rinfo = replicate_all_target_info(&roi, &target);
+
+        let action = resolve_replicate_all_action(
+            ReplicateAllActionContext {
+                roi: &roi,
+                tgt_client: &target,
+                bucket: &roi.bucket,
+                object: &roi.name,
+                start_time: OffsetDateTime::now_utc(),
+                ssec_audit_required: false,
+            },
+            object_info,
+            &mut rinfo,
+        )
+        .await;
+
+        assert!(matches!(action, Some((ReplicationAction::All, _))));
+        assert!(rinfo.error.is_none());
+        server.join().expect("test HTTP server should finish");
+    }
+
     async fn register_test_target(target: &Arc<TargetClient>) {
         ReplicationTargetStore::register_test_target(target).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()
@@ -663,20 +672,39 @@ 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/ops/multipart.rs

@@ -2318,6 +2318,43 @@ 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. Only an absent version has no local state to compare;
+        // other read failures must leave the upload retryable rather than
+        // committing inbound metadata without the LWW check.
+        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) => return Err(err),
+            }
+        }
+
         for meta in parts_metadatas.iter_mut() {
             if meta.has_valid_erasure_geometry() {
                 meta.size = fi.size;
@@ -7055,6 +7092,146 @@ 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()
+            };
+
+            // Make the destination version unreadable on quorum while the
+            // staged upload remains intact. The commit barrier lets the old
+            // fail-open path move past the LWW read; restoring the metadata
+            // there proves it would otherwise commit the stale tags.
+            let mut damaged_metadata = Vec::new();
+            for temp_dir in temp_dirs.iter().take(3) {
+                let path = temp_dir.path().join(bucket).join(object).join(STORAGE_FORMAT_FILE);
+                let original = tokio::fs::read(&path).await.expect("destination xl.meta should be readable");
+                tokio::fs::write(&path, b"not an xl.meta")
+                    .await
+                    .expect("destination xl.meta should be corruptible");
+                damaged_metadata.push((path, original));
+            }
+
+            let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::BeforeLockLost);
+            let first_set = set_disks.clone();
+            let first_upload_id = upload_id.clone();
+            let first_parts = parts.clone();
+            let first_opts = complete_opts.clone();
+            let mut first_completion = tokio::spawn(async move {
+                first_set
+                    .complete_multipart_upload(bucket, object, &first_upload_id, first_parts, &first_opts)
+                    .await
+            });
+
+            let first_result = tokio::select! {
+                result = &mut first_completion => result.expect("first completion task should finish"),
+                () = barrier.wait_until_paused() => {
+                    for (path, original) in &damaged_metadata {
+                        tokio::fs::write(path, original).await.expect("destination xl.meta should be restorable");
+                    }
+                    barrier.release();
+                    first_completion.await.expect("released completion task should finish")
+                }
+            };
+            for (path, original) in &damaged_metadata {
+                tokio::fs::write(path, original)
+                    .await
+                    .expect("destination xl.meta should be restored");
+            }
+            drop(barrier);
+
+            let first_error = first_result.expect_err("unreadable destination metadata must fail before multipart commit");
+            assert!(
+                !(is_err_object_not_found(&first_error) || is_err_version_not_found(&first_error)),
+                "corrupt destination metadata must not be treated as an absent version: {first_error}"
+            );
+
+            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

@@ -1881,6 +1881,110 @@ 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,
@@ -2073,6 +2177,14 @@ impl SetDisks {
                 user_defined.insert(key.clone(), value.clone());
             }
         }
+        if replication_lww_applicable(opts) {
+            // Object Lock evaluation stamps category timestamps with this
+            // receiver's clock. Pin them back to the source-authored times
+            // before the first copy of a version is committed; the existing-
+            // version branch below may still replace them with newer local
+            // state.
+            merge_replication_metadata_lww(&mut user_defined, &HashMap::new(), opts);
+        }
         if expected_restore_operation_id.is_some() {
             rustfs_utils::http::metadata_compat::remove_str(&mut user_defined, SUFFIX_RESTORE_OPERATION_ID);
         }
@@ -2562,6 +2674,22 @@ impl SetDisks {
                         if check_object_lock_for_deletion_with_state(object_lock_config.state(), &existing, false)?.is_some() {
                             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),
@@ -8082,6 +8210,387 @@ 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 first_inbound_version_pins_source_authored_timestamp() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "lww-first-version-ts-pinned";
+        let object = "object";
+        let version_id = Uuid::new_v4().to_string();
+        make_bucket(&disk_stores, bucket).await;
+
+        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, T_OLD.to_string());
+        let mut evaluated = inbound.clone();
+        insert_str(&mut evaluated, SUFFIX_OBJECTLOCK_LEGALHOLD_TIMESTAMP, T_NEW.to_string());
+        let opts = ObjectOptions {
+            replication_request: true,
+            replication_legalhold_timestamp: Some(parse_ts(T_OLD)),
+            eval_metadata: Some(evaluated),
+            ..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_LEGALHOLD_TIMESTAMP).as_deref(),
+            Some(T_OLD),
+            "the first copy must store the source timestamp, not the receiver evaluation time"
+        );
+    }
+
+    #[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));
+    }
+}
+
 #[cfg(test)]
 mod inline_put_commit_path_tests {
     use super::hermetic_set_disks_support::hermetic_set_disks_isolated as hermetic_set_disks;

crates/replication/src/operation.rs

@@ -88,6 +88,17 @@ impl MustReplicateOptions {
             return true;
         }
 
+        // A REPLICA version was delivered by a peer and carries no per-target
+        // internal status of its own; whether its local metadata edits flow
+        // back is the replication rule's ReplicaModifications decision
+        // (`ReplicationConfig::replicate` with `replica = true`, MinIO
+        // mustReplicate parity). Gating it on a COMPLETED target state would
+        // silently drop every replica-side tag / retention / legal-hold edit in
+        // an active-active topology (rustfs/backlog#1953).
+        if self.replication_status() == ReplicationStatusType::Replica {
+            return true;
+        }
+
         get_internal_metadata(&self.meta, SUFFIX_REPLICATION_STATUS)
             .as_deref()
             .and_then(|statuses| {
@@ -381,6 +392,13 @@ mod tests {
 
         assert!(options.metadata_target_is_eligible(arn));
         assert!(!options.metadata_target_is_eligible("arn:rustfs:replication:missing"));
+
+        // A replica-side metadata edit (active-active) has no per-target
+        // internal status; eligibility is left to the ReplicaModifications rule.
+        let replica = MustReplicateOptions::new(&HashMap::new(), String::new(), ReplicationType::Metadata, false)
+            .with_replication_status(ReplicationStatusType::Replica);
+        assert!(replica.metadata_target_is_eligible(arn));
+        assert!(replica.metadata_target_is_eligible("arn:rustfs:replication:missing"));
         assert!(
             MustReplicateOptions::new(&HashMap::new(), String::new(), ReplicationType::Object, false)
                 .metadata_target_is_eligible(arn)

crates/utils/src/http/headers.rs

@@ -47,6 +47,8 @@ pub const AMZ_DELETE_MARKER: &str = "x-amz-delete-marker";
 
 // S3 object tagging
 pub const AMZ_OBJECT_TAGGING: &str = "X-Amz-Tagging";
+/// Lowercase wire form of [`AMZ_OBJECT_TAGGING`] for `HeaderMap` insertion.
+pub const AMZ_OBJECT_TAGGING_LOWER: &str = "x-amz-tagging";
 pub const AMZ_TAG_COUNT: &str = "x-amz-tagging-count";
 pub const AMZ_TAG_DIRECTIVE: &str = "X-Amz-Tagging-Directive";