ci: model server as a composition root

Swift container and account metadata handlers cloned the cached
BucketMetadata, set the tagging fields, and called set_bucket_metadata,
which only updates the in-memory cache map. Nothing reached
.metadata.bin, so every Swift metadata POST was lost on restart and
silently overwritten by the next disk-truth reload (a peer
LoadBucketMetadata notification or the 15-minute refresh loop) — while
the client had already been told 2xx.

Route these writes through a new metadata_sys::update_config_with: a
read-modify-write that loads the on-disk metadata and persists the
result under the same write guard metadata_sys::update uses, so the
rewrite merges against disk truth instead of a possibly stale cache and
cannot clobber a concurrent update to another config file. Peers are
notified afterwards, matching the S3 config handlers.

Persisting these writes required hardening the paths that now produce
durable state:

- Account metadata writes validate account ownership. This metadata
  holds the account's TempURL signing key, so an unauthenticated write
  for someone else's account would have become a durable, cluster-wide
  takeover of that account's pre-signed URLs. Reads stay open because
  TempURL signature validation runs before credentials exist.
- disable_versioning verifies the container exists. Without it the
  metadata loader's "no metadata on disk" default would be persisted,
  creating an orphan metadata file and caching a fabricated default as
  authoritative.
- Container and account metadata are size- and count-limited, reusing
  the Swift limits object metadata already enforces; these tags land in
  the bucket metadata file that every later config write rewrites whole.
- A rewrite refuses to run when the persisted tagging config is
  unreadable, instead of merging onto an empty set and wiping the
  container ACL and versioning tags. It reports 409 naming the remedy.
- Storage errors are logged in full and reported generically, since they
  now carry real disk and quorum detail.

The tagging arm of BucketMetadata::update_config also clears the parsed
config, as the lifecycle arm does: parse_all_configs skips empty XML
rather than clearing, so a cleared config kept serving the old tags.
Tagging is serialized with the S3 XML serializer the loader can parse
back, not quick_xml, whose output was never round-trippable.

.agents/skills/arch-checks/SKILL.md

@@ -10,10 +10,16 @@ never weaken a check to get green.
 
 ## `check_layer_dependencies.sh` — layer DAG in `rustfs/src`
 
-Enforces `interface (admin, storage/ecfs, storage/s3_api) → app → infra`; no
-upward imports. Known legacy violations live in
+Enforces `composition (server, startup/init) → interface (admin,
+storage/ecfs, storage/s3_api) → app → infra`; no upward imports. Server source
+files are composition roots, while imports of their exported HTTP contracts
+are classified as interface dependencies. Known legacy violations live in
 `scripts/layer-dependency-baseline.txt`.
 
+Dedicated `*_test.rs` and `tests/` modules are outside this production guard.
+Inline `#[cfg(test)]` imports remain checked under their source file's layer;
+move architecture-crossing test scaffolding into a dedicated test module.
+
 - **New violation**: restructure your change so the dependency points
   downward (move the shared type/function to the lower layer).
 - **You legitimately removed a baseline entry**: run

.config/nextest.toml

@@ -1,17 +1,14 @@
 # nextest configuration for RustFS.
 #
-# Serialize two known load-sensitive / global-state-sharing ecstore test groups
-# so the full parallel nextest suite stops producing spurious failures
-# (backlog #937). These tests pass in isolation but flake under the loaded
-# parallel run for two distinct reasons:
+# Serialize the ecstore tests that share the process-wide disk registry or
+# exercise a multi-disk commit handoff across nextest process boundaries.
 #
 #   * store::bucket::tests::bucket_delete_* share process/global state (disk
 #     registry, lock client) and race make_bucket into InsufficientWriteQuorum
 #     when run concurrently with other ecstore tests.
 #   * bucket_lifecycle_ops::tests::concurrent_resend_same_part_commits_one_generation
-#     asserts a lock-acquire correctness property whose serialized cross-disk
-#     commits exceed the (already max'd, 60s) acquire deadline only when the
-#     suite saturates disk I/O.
+#     uses the shared multipart fixture and a deterministic uploadId-lock
+#     handoff, so it must not overlap another process mutating that fixture.
 #
 # serial_test's #[serial] attribute does NOT serialize these across runs:
 # nextest executes each test in its own process, where the in-process
@@ -100,13 +97,6 @@ path = "junit.xml"
 # profile's own overrides list, not the default profile's).
 # ===========================================================================
 
-# QUARANTINE: OPEN backlog#937 — concurrent_resend lock-acquire deadline flakes
-# under saturated disk I/O in the full parallel suite.
-[[profile.ci.overrides]]
-filter = 'package(rustfs-ecstore) & test(concurrent_resend_same_part_commits_one_generation)'
-test-group = 'ecstore-serial-flaky'
-retries = 2
-
 # QUARANTINE: OPEN backlog#937 — store::bucket::tests::bucket_delete_* race
 # make_bucket into InsufficientWriteQuorum via shared global state under load.
 [[profile.ci.overrides]]
@@ -114,6 +104,11 @@ filter = 'package(rustfs-ecstore) & test(/^store::bucket::tests::bucket_delete_(
 test-group = 'ecstore-serial-flaky'
 retries = 2
 
+# Keep the deterministic multipart handoff isolated across nextest processes.
+[[profile.ci.overrides]]
+filter = 'package(rustfs-ecstore) & test(concurrent_resend_same_part_commits_one_generation)'
+test-group = 'ecstore-serial-flaky'
+
 # QUARANTINE: OPEN rustfs#4690 — walk_dir stall-budget accounting test depends
 # on producer/consumer timing windows that stretch past the budget on loaded
 # CI runners (regression test for rustfs#4644; failed on a zero-Rust-diff PR).

Cargo.lock

@@ -9732,6 +9732,7 @@ dependencies = [
  "rustfs-policy",
  "rustfs-rio",
  "rustfs-storage-api",
+ "rustfs-test-utils",
  "rustfs-tls-runtime",
  "rustfs-trusted-proxies",
  "rustfs-utils",

crates/ecstore/src/api/mod.rs

@@ -128,7 +128,7 @@ pub mod bucket {
             get_object_lock_config, get_public_access_block_config, get_quota_config, get_replication_config,
             get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config, get_website_config,
             init_bucket_metadata_sys, list_bucket_targets, remove_bucket_metadata, set_bucket_metadata, update,
-            update_bucket_targets_under_transaction_lock,
+            update_bucket_targets_under_transaction_lock, update_config_with,
         };
     }
 

crates/ecstore/src/bucket/lifecycle/bucket_lifecycle_ops.rs

@@ -11112,6 +11112,7 @@ mod tests {
     #[tokio::test(flavor = "multi_thread")]
     #[serial]
     async fn concurrent_resend_same_part_commits_one_generation() {
+        use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause};
         use crate::storage_api_contracts::object::ObjectIO as _;
 
         let (_paths, ecstore) = setup_test_env().await;
@@ -11133,51 +11134,36 @@ mod tests {
             })
             .collect();
 
-        // Two independent causes can produce a spurious lock-acquire timeout
-        // here, and both must stay covered:
-        // 1. A lost/stolen fast-lock wakeup could strand a waiter until the
-        //    deadline — fixed for real in fast_lock::shard by bounding each
-        //    notification wait (NOTIFY_WAIT_CAP re-polling).
-        // 2. Under the full nextest suite on loaded CI disks, the
-        //    *legitimately serialized* cross-disk commits can exceed the
-        //    acquire deadline all by themselves — observed on CI at the 5s
-        //    default and the 30s production default with six resends, and
-        //    again at 60s, which is a hard ceiling: fast_lock clamps every
-        //    requested timeout to MAX_ACQUIRE_TIMEOUT (60s), so raising the
-        //    env override higher is a no-op (the Timeout error still reports
-        //    the requested value). Keep the guard about the correctness
-        //    property, not disk latency: request the full 60s ceiling and cap
-        //    the queue depth at three resends, so the last waiter sits behind
-        //    at most two serialized commits (~12s each on the slowest observed
-        //    CI runner, comfortably inside the deadline). Three concurrent
-        //    resends still race the streaming phase and contend on the commit
-        //    lock, which is all the generation-mixing regression needs.
-        //    `#[serial]` keeps the process-wide env override isolated.
-        let results = temp_env::async_with_vars([(rustfs_config::ENV_OBJECT_LOCK_ACQUIRE_TIMEOUT, Some("60"))], async {
-            let mut tasks = tokio::task::JoinSet::new();
-            for payload in candidates.iter().cloned() {
-                let store = ecstore.clone();
-                let bucket = bucket.clone();
-                let upload_id = upload.upload_id.clone();
-                tasks.spawn(async move {
-                    let mut data = PutObjReader::from_vec(payload.clone());
-                    store
-                        .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
-                        .await
-                        .map(|info| (info, payload))
-                });
-            }
+        let commit_barrier = MultipartCommitBarrier::install(&bucket, object, MultipartCommitPause::PutPartBeforeLockLost);
+        let start = Arc::new(tokio::sync::Barrier::new(candidates.len() + 1));
+        let mut tasks = tokio::task::JoinSet::new();
+        for payload in candidates.iter().cloned() {
+            let store = ecstore.clone();
+            let bucket = bucket.clone();
+            let upload_id = upload.upload_id.clone();
+            let start = Arc::clone(&start);
+            tasks.spawn(async move {
+                start.wait().await;
+                let mut data = PutObjReader::from_vec(payload.clone());
+                store
+                    .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
+                    .await
+                    .map(|info| (info, payload))
+            });
+        }
+        start.wait().await;
 
-            // Every concurrent resend must succeed; the commit lock must never
-            // starve a waiter into a timeout.
-            let mut results = Vec::new();
-            while let Some(joined) = tasks.join_next().await {
-                let outcome = joined.expect("put_object_part task should not panic");
-                results.push(outcome.expect("every concurrent same-part resend must succeed without lock timeout"));
-            }
-            results
-        })
-        .await;
+        // The first writer holds the uploadId commit lock while the other
+        // resends reach the same critical section. Releasing it proves the
+        // handoff without depending on saturated CI disk latency.
+        commit_barrier.wait_until_paused().await;
+        commit_barrier.release();
+
+        let mut results = Vec::new();
+        while let Some(joined) = tasks.join_next().await {
+            let outcome = joined.expect("put_object_part task should not panic");
+            results.push(outcome.expect("every concurrent same-part resend must succeed without lock timeout"));
+        }
         assert_eq!(results.len(), candidates.len());
 
         // Exactly one generation is visible after the serialized commits, and its

crates/ecstore/src/bucket/metadata.rs

@@ -737,6 +737,9 @@ impl BucketMetadata {
             }
             BUCKET_TAGGING_CONFIG => {
                 self.tagging_config_xml = data;
+                // Drop the parsed form (like lifecycle above) so clearing the
+                // payload can't leave stale parsed tags to be cached.
+                self.tagging_config = None;
                 self.tagging_config_updated_at = updated;
             }
             BUCKET_QUOTA_CONFIG_FILE => {
@@ -1318,6 +1321,30 @@ mod test {
         assert!(bm.lifecycle_config.is_none());
     }
 
+    /// Companion to the lifecycle case above. `parse_all_configs` skips empty
+    /// XML rather than clearing, so without the explicit reset a cleared
+    /// tagging config would keep serving the previously parsed tags.
+    #[test]
+    fn tagging_update_config_clears_parsed_config_on_delete() {
+        let mut bm = BucketMetadata::new("test-bucket");
+        let tagging_xml = br#"<Tagging><TagSet><Tag><Key>env</Key><Value>prod</Value></Tag></TagSet></Tagging>"#;
+
+        bm.update_config(BUCKET_TAGGING_CONFIG, tagging_xml.to_vec())
+            .expect("tagging config should update");
+        bm.parse_all_configs().expect("tagging config should parse");
+        assert!(bm.tagging_config.is_some());
+
+        bm.update_config(BUCKET_TAGGING_CONFIG, Vec::new())
+            .expect("tagging config delete should update metadata");
+
+        assert!(bm.tagging_config_xml.is_empty());
+        assert!(bm.tagging_config.is_none());
+
+        // A re-parse must not resurrect them either.
+        bm.parse_all_configs().expect("cleared tagging should parse");
+        assert!(bm.tagging_config.is_none());
+    }
+
     #[tokio::test]
     async fn marshal_msg_complete_example() {
         // Create a complete BucketMetadata with various configurations

crates/ecstore/src/bucket/metadata_sys.rs

@@ -239,6 +239,35 @@ pub async fn update_bucket_targets_under_transaction_lock(bucket: &str, data: Ve
     bucket_meta_sys.update(bucket, BUCKET_TARGETS_FILE, data).await
 }
 
+/// Read-modify-write one bucket config file under the metadata system's
+/// outer write guard.
+///
+/// `mutate` sees the freshly loaded on-disk metadata and returns the
+/// replacement payload for `config_file` (empty clears it, like
+/// [`delete`]). Both the read and the persisted write happen inside the
+/// same guard that [`update`] uses, so within this process the rewrite can
+/// neither clobber a concurrent update to another config file nor lose a
+/// concurrent write to the same one — unlike caching a mutated clone of
+/// previously read metadata.
+///
+/// This guard is process-local. Writers on other nodes still race, exactly
+/// as they do for [`update`]: each rewrites the whole metadata file, so the
+/// later save wins. What this narrows is the window — from "as stale as the
+/// local cache" down to a single metadata read plus write.
+pub async fn update_config_with<F>(bucket: &str, config_file: &str, mutate: F) -> Result<OffsetDateTime>
+where
+    F: FnOnce(&BucketMetadata) -> Result<Vec<u8>> + Send,
+{
+    let bucket_meta_sys_lock = get_bucket_metadata_sys()?;
+    let _targets_guard = if config_file == BUCKET_TARGETS_FILE {
+        Some(acquire_bucket_targets_transaction_lock(bucket).await?)
+    } else {
+        None
+    };
+    let mut bucket_meta_sys = bucket_meta_sys_lock.write().await;
+    bucket_meta_sys.update_config_with(bucket, config_file, mutate).await
+}
+
 pub async fn acquire_bucket_targets_transaction_lock(bucket: &str) -> Result<rustfs_lock::NamespaceLockGuard> {
     let bucket_meta_sys_lock = get_bucket_metadata_sys()?;
     let api = bucket_meta_sys_lock.read().await.object_store();
@@ -603,24 +632,7 @@ impl BucketMetadataSys {
             return Err(Error::other("errServerNotInitialized"));
         };
 
-        if is_meta_bucketname(bucket) {
-            return Err(Error::other("errInvalidArgument"));
-        }
-
-        let mut bm = match load_bucket_metadata_parse(store, bucket, parse).await {
-            Ok(res) => res,
-            Err(err) => {
-                if !runtime_sources::setup_is_erasure().await
-                    && !runtime_sources::setup_is_dist_erasure().await
-                    && is_err_bucket_not_found(&err)
-                {
-                    BucketMetadata::new(bucket)
-                } else {
-                    error!("load bucket metadata failed: {}", err);
-                    return Err(err);
-                }
-            }
-        };
+        let mut bm = Self::load_bucket_metadata_for_update(store, bucket, parse).await?;
 
         let updated = bm.update_config(config_file, data)?;
 
@@ -629,6 +641,49 @@ impl BucketMetadataSys {
         Ok(updated)
     }
 
+    /// See the free [`update_config_with`]: same load-mutate-persist cycle as
+    /// [`Self::update`], with the payload computed from the loaded metadata
+    /// instead of supplied up front. Loads through this system's own store so
+    /// the read and the persisted write target the same instance.
+    async fn update_config_with<F>(&mut self, bucket: &str, config_file: &str, mutate: F) -> Result<OffsetDateTime>
+    where
+        F: FnOnce(&BucketMetadata) -> Result<Vec<u8>> + Send,
+    {
+        let mut bm = Self::load_bucket_metadata_for_update(self.api.clone(), bucket, true).await?;
+
+        let data = mutate(&bm)?;
+        let updated = bm.update_config(config_file, data)?;
+
+        self.save(bm).await?;
+
+        Ok(updated)
+    }
+
+    /// Load a bucket's on-disk metadata as the base of a config rewrite.
+    /// Outside erasure setups a missing metadata file degrades to a fresh
+    /// default (legacy buckets without one); erasure setups fail instead of
+    /// fabricating state that a quorum may still hold.
+    async fn load_bucket_metadata_for_update(store: Arc<ECStore>, bucket: &str, parse: bool) -> Result<BucketMetadata> {
+        if is_meta_bucketname(bucket) {
+            return Err(Error::other("errInvalidArgument"));
+        }
+
+        match load_bucket_metadata_parse(store, bucket, parse).await {
+            Ok(res) => Ok(res),
+            Err(err) => {
+                if !runtime_sources::setup_is_erasure().await
+                    && !runtime_sources::setup_is_dist_erasure().await
+                    && is_err_bucket_not_found(&err)
+                {
+                    Ok(BucketMetadata::new(bucket))
+                } else {
+                    error!("load bucket metadata failed: {}", err);
+                    Err(err)
+                }
+            }
+        }
+    }
+
     async fn save(&self, bm: BucketMetadata) -> Result<()> {
         if is_meta_bucketname(&bm.name) {
             return Err(Error::other("errInvalidArgument"));
@@ -1068,6 +1123,122 @@ mod tests {
 
         assert!(matches!(err, Error::Io(_)), "malformed persisted policy must surface its parse failure");
     }
+    /// A tagging rewrite through `update_config_with` (the Swift metadata
+    /// POST path) is persisted: it survives a metadata reload from disk, and
+    /// an emptied rewrite clears the config in the cached copy too instead of
+    /// leaving stale parsed tags behind.
+    #[tokio::test]
+    async fn update_config_with_persists_tagging_rewrite_across_disk_reload() {
+        use crate::bucket::metadata::BUCKET_TAGGING_CONFIG;
+        use s3s::dto::Tag;
+
+        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let mut sys = BucketMetadataSys::new(ecstore);
+
+        let bucket = "swift-tagging-bucket";
+        sys.persist_and_set(BucketMetadata::new(bucket))
+            .await
+            .expect("initial metadata should persist");
+
+        let tagging = Tagging {
+            tag_set: vec![Tag {
+                key: Some("swift-meta-color".to_string()),
+                value: Some("blue".to_string()),
+            }],
+        };
+        let xml = crate::bucket::utils::serialize::<Tagging>(&tagging).expect("tagging should serialize");
+        sys.update_config_with(bucket, BUCKET_TAGGING_CONFIG, move |bm| {
+            assert!(bm.tagging_config.is_none(), "rewrite must see the on-disk state");
+            Ok(xml)
+        })
+        .await
+        .expect("tagging rewrite should persist");
+
+        // Simulate the disk-truth reload that used to lose Swift writes: drop
+        // the cached entry and lazily re-load from the metadata file.
+        sys.metadata_map.write().await.clear();
+        let (tags, _) = sys
+            .get_tagging_config(bucket)
+            .await
+            .expect("tagging must survive a reload from disk");
+        assert_eq!(tags.tag_set.len(), 1);
+        assert_eq!(tags.tag_set[0].key.as_deref(), Some("swift-meta-color"));
+        assert_eq!(tags.tag_set[0].value.as_deref(), Some("blue"));
+
+        // An emptied rewrite clears the config everywhere.
+        sys.update_config_with(bucket, BUCKET_TAGGING_CONFIG, |bm| {
+            assert!(bm.tagging_config.is_some(), "rewrite must see the persisted tags");
+            Ok(Vec::new())
+        })
+        .await
+        .expect("clearing rewrite should persist");
+        assert_eq!(
+            sys.get_tagging_config(bucket).await.unwrap_err(),
+            Error::ConfigNotFound,
+            "cleared tagging must not be served from the cache"
+        );
+        sys.metadata_map.write().await.clear();
+        assert_eq!(
+            sys.get_tagging_config(bucket).await.unwrap_err(),
+            Error::ConfigNotFound,
+            "cleared tagging must not reappear after a reload from disk"
+        );
+    }
+
+    /// The load and the persisted write share one write guard, so concurrent
+    /// rewrites of the same config compose instead of clobbering each other.
+    /// Moving the load outside that guard loses all but the last tag.
+    #[tokio::test]
+    async fn concurrent_update_config_with_calls_do_not_lose_writes() {
+        use crate::bucket::metadata::BUCKET_TAGGING_CONFIG;
+        use s3s::dto::Tag;
+
+        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(ecstore)));
+
+        let bucket = "swift-tagging-concurrent";
+        sys.read()
+            .await
+            .persist_and_set(BucketMetadata::new(bucket))
+            .await
+            .expect("initial metadata should persist");
+
+        const WRITERS: usize = 8;
+        let mut handles = Vec::with_capacity(WRITERS);
+        for idx in 0..WRITERS {
+            let sys = sys.clone();
+            handles.push(tokio::spawn(async move {
+                sys.write()
+                    .await
+                    .update_config_with(bucket, BUCKET_TAGGING_CONFIG, move |bm| {
+                        // Each writer merges its own tag onto whatever is
+                        // currently persisted — the Swift rewrite shape.
+                        let mut tagging = bm.tagging_config.clone().unwrap_or_else(|| Tagging { tag_set: vec![] });
+                        tagging.tag_set.push(Tag {
+                            key: Some(format!("swift-meta-key{idx}")),
+                            value: Some(idx.to_string()),
+                        });
+                        crate::bucket::utils::serialize::<Tagging>(&tagging).map_err(|e| Error::other(e.to_string()))
+                    })
+                    .await
+            }));
+        }
+
+        for handle in handles {
+            handle
+                .await
+                .expect("writer task should join")
+                .expect("rewrite should persist");
+        }
+
+        let (tags, _) = sys
+            .read()
+            .await
+            .get_tagging_config(bucket)
+            .await
+            .expect("tagging should be readable");
+        assert_eq!(tags.tag_set.len(), WRITERS, "every concurrent rewrite must survive: {tags:?}");
+    }
 
     fn target(bucket: &str, id: &str) -> BucketTarget {
         BucketTarget {

crates/ecstore/src/set_disk/mod.rs

@@ -687,6 +687,8 @@ mod core;
 mod ctx;
 mod metadata;
 mod ops;
+#[cfg(test)]
+pub(crate) use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
 #[cfg(feature = "test-util")]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
 pub(crate) use ops::object::body_cache_plaintext_len;

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

@@ -29,6 +29,12 @@ impl SetDisks {
     pub async fn delete_all(&self, bucket: &str, prefix: &str) -> Result<()> {
         ListOperations::new(self.ctx()).delete_all(bucket, prefix).await
     }
+
+    pub(crate) async fn delete_all_with_quorum(&self, bucket: &str, prefix: &str, write_quorum: usize) -> Result<()> {
+        ListOperations::new(self.ctx())
+            .delete_all_with_quorum(bucket, prefix, write_quorum)
+            .await
+    }
 }
 
 /// List/prefix maintenance operations, borrowing the `SetDisks` core state
@@ -48,6 +54,14 @@ impl<'a> ListOperations<'a> {
     }
 
     pub(crate) async fn delete_all(&self, bucket: &str, prefix: &str) -> Result<()> {
+        self.delete_all_inner(bucket, prefix, None).await
+    }
+
+    async fn delete_all_with_quorum(&self, bucket: &str, prefix: &str, write_quorum: usize) -> Result<()> {
+        self.delete_all_inner(bucket, prefix, Some(write_quorum)).await
+    }
+
+    async fn delete_all_inner(&self, bucket: &str, prefix: &str, write_quorum: Option<usize>) -> Result<()> {
         let disks = self.ctx.disks().read().await;
 
         let disks = disks.clone();
@@ -79,6 +93,9 @@ impl<'a> ListOperations<'a> {
                 Ok(_) => {
                     errors.push(None);
                 }
+                Err(DiskError::FileNotFound | DiskError::PathNotFound | DiskError::VolumeNotFound) => {
+                    errors.push(None);
+                }
                 Err(e) => {
                     errors.push(Some(e));
                 }
@@ -97,6 +114,12 @@ impl<'a> ListOperations<'a> {
             );
         }
 
+        if let Some(write_quorum) = write_quorum
+            && let Some(err) = reduce_write_quorum_errs(&errors, OBJECT_OP_IGNORED_ERRS, write_quorum)
+        {
+            return Err(err.into());
+        }
+
         Ok(())
     }
 }

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

@@ -26,6 +26,8 @@ use crate::crash_inject::{self, CrashPoint};
 use crate::multipart_listing::paginate_multipart_listing;
 use futures::{StreamExt, stream};
 use std::future::Future;
+#[cfg(test)]
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 use tokio::task::JoinSet;
 
@@ -33,7 +35,7 @@ const MULTIPART_LIST_IO_CONCURRENCY: usize = 16;
 
 #[cfg(test)]
 #[derive(Clone, Copy, PartialEq, Eq)]
-enum MultipartCommitPause {
+pub(crate) enum MultipartCommitPause {
     PutPartBeforeLockLost,
     PutPartAfterRename,
     BeforeLockLost,
@@ -45,12 +47,13 @@ struct MultipartCommitBarrierState {
     bucket: String,
     object: String,
     pause: MultipartCommitPause,
+    armed: AtomicBool,
     arrived: tokio::sync::Notify,
     release: tokio::sync::Notify,
 }
 
 #[cfg(test)]
-struct MultipartCommitBarrier {
+pub(crate) struct MultipartCommitBarrier {
     state: Arc<MultipartCommitBarrierState>,
 }
 
@@ -60,11 +63,12 @@ static MULTIPART_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc
 
 #[cfg(test)]
 impl MultipartCommitBarrier {
-    fn install(bucket: &str, object: &str, pause: MultipartCommitPause) -> Self {
+    pub(crate) fn install(bucket: &str, object: &str, pause: MultipartCommitPause) -> Self {
         let state = Arc::new(MultipartCommitBarrierState {
             bucket: bucket.to_string(),
             object: object.to_string(),
             pause,
+            armed: AtomicBool::new(true),
             arrived: tokio::sync::Notify::new(),
             release: tokio::sync::Notify::new(),
         });
@@ -78,13 +82,13 @@ impl MultipartCommitBarrier {
         Self { state }
     }
 
-    async fn wait_until_paused(&self) {
+    pub(crate) async fn wait_until_paused(&self) {
         tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
             .await
             .expect("multipart completion should reach the deterministic commit barrier");
     }
 
-    fn release(&self) {
+    pub(crate) fn release(&self) {
         self.state.release.notify_one();
     }
 }
@@ -112,7 +116,9 @@ async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartComm
         .as_ref()
         .filter(|barrier| barrier.bucket == bucket && barrier.object == object && barrier.pause == pause)
         .cloned();
-    if let Some(barrier) = barrier {
+    if let Some(barrier) = barrier
+        && barrier.armed.swap(false, Ordering::AcqRel)
+    {
         barrier.arrived.notify_one();
         barrier.release.notified().await;
     }
@@ -777,6 +783,9 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
         mut max_parts: usize,
         opts: &ObjectOptions,
     ) -> Result<ListPartsInfo> {
+        let _upload_guard = self
+            .acquire_multipart_upload_read_lock("list_object_parts", bucket, object, upload_id, opts)
+            .await?;
         let (fi, _) = self.check_upload_id_exists(bucket, object, upload_id, false).await?;
 
         let upload_id_path = Self::get_upload_id_dir(bucket, object, upload_id);
@@ -1254,10 +1263,15 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
         let _upload_guard = self
             .acquire_multipart_upload_write_lock("abort_multipart_upload", bucket, object, upload_id, opts)
             .await?;
-        self.check_upload_id_exists(bucket, object, upload_id, false).await?;
+        let (fi, _) = self.check_upload_id_exists(bucket, object, upload_id, true).await?;
         let upload_id_path = Self::get_upload_id_dir(bucket, object, upload_id);
 
-        self.delete_all(RUSTFS_META_MULTIPART_BUCKET, &upload_id_path).await
+        self.delete_all_with_quorum(
+            RUSTFS_META_MULTIPART_BUCKET,
+            &upload_id_path,
+            fi.write_quorum(self.default_write_quorum()),
+        )
+        .await
     }
     // complete_multipart_upload finished
     #[tracing::instrument(skip(self))]
@@ -1815,7 +1829,36 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
 
         #[cfg(test)]
         pause_multipart_commit(bucket, object, MultipartCommitPause::AfterRename).await;
-        drop(upload_guard);
+
+        let cleanup_store = self.clone();
+        let cleanup_upload_id_path = upload_id_path.clone();
+        let cleanup_bucket = bucket.to_owned();
+        let cleanup_object = object.to_owned();
+        let cleanup_upload_id = upload_id.to_owned();
+        let cleanup_handle = tokio::spawn(async move {
+            let _upload_guard = upload_guard;
+            if let Err(err) = cleanup_store
+                .delete_all_with_quorum(RUSTFS_META_MULTIPART_BUCKET, &cleanup_upload_id_path, write_quorum)
+                .await
+            {
+                warn!(
+                    bucket = %cleanup_bucket,
+                    object = %cleanup_object,
+                    upload_id = %cleanup_upload_id,
+                    error = ?err,
+                    "completed multipart upload staging cleanup did not reach write quorum"
+                );
+            }
+        });
+        if let Err(err) = cleanup_handle.await {
+            warn!(
+                bucket = %bucket,
+                object = %object,
+                upload_id = %upload_id,
+                error = ?err,
+                "completed multipart upload staging cleanup task failed"
+            );
+        }
         drop(object_lock_guard); // drop object lock guard to release the lock
 
         // backlog#1321: enqueue heal only when the committed replicas actually
@@ -1860,12 +1903,6 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
             });
         }
 
-        let upload_id_path = upload_id_path.clone();
-        let store = self.clone();
-        let _cleanup_handle = tokio::spawn(async move {
-            let _ = store.delete_all(RUSTFS_META_MULTIPART_BUCKET, &upload_id_path).await;
-        });
-
         for (i, op_disk) in online_disks.iter().enumerate() {
             if let Some(disk) = op_disk
                 && disk.is_online().await
@@ -2177,6 +2214,122 @@ mod tests {
         )
     }
 
+    async fn assert_complete_first_linearizes(bucket: &'static str, object: &'static str, create_opts: ObjectOptions) {
+        let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+        let signaling = Arc::new(SignalingLockClient::new(Arc::new(LocalClient::with_manager(manager))));
+        let lockers: Vec<Arc<dyn LockClient>> = vec![signaling.clone()];
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks_with_lockers(4, 0, 2, lockers).await;
+        make_bucket_on_all(&disk_stores, bucket).await;
+        let (upload_id, parts) = stage_upload_with_create_opts(&set_disks, bucket, object, &[0x47; 4096], &create_opts).await;
+        let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        signaling.set_target(rustfs_lock::ObjectKey::new(RUSTFS_META_MULTIPART_BUCKET, upload_id_path));
+        let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;
+        let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::AfterRename);
+
+        let complete_store = set_disks.clone();
+        let complete_upload_id = upload_id.clone();
+        let complete = tokio::spawn(async move {
+            complete_store
+                .complete_multipart_upload(bucket, object, &complete_upload_id, parts, &ObjectOptions::default())
+                .await
+        });
+        barrier.wait_until_paused().await;
+
+        let abort_store = set_disks.clone();
+        let abort_upload_id = upload_id.clone();
+        let abort = tokio::spawn(async move {
+            abort_store
+                .abort_multipart_upload(bucket, object, &abort_upload_id, &ObjectOptions::default())
+                .await
+        });
+        signaling.wait_for_attempts(2).await;
+        assert!(!abort.is_finished(), "abort must wait for the completion upload lock");
+
+        barrier.release();
+        complete
+            .await
+            .expect("completion task should not panic")
+            .expect("completion should win the upload finalization");
+        let abort_err = abort
+            .await
+            .expect("abort task should not panic")
+            .expect_err("abort must observe the upload as finalized");
+        assert!(matches!(abort_err, StorageError::InvalidUploadID(..)));
+        set_disks
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("complete-first must leave the committed object readable");
+        assert!(matches!(
+            set_disks.check_upload_id_exists(bucket, object, &upload_id, false).await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+    }
+
+    async fn assert_abort_first_linearizes(bucket: &'static str, object: &'static str, create_opts: ObjectOptions) {
+        let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+        let signaling = Arc::new(SignalingLockClient::new(Arc::new(LocalClient::with_manager(manager))));
+        let lockers: Vec<Arc<dyn LockClient>> = vec![signaling.clone()];
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks_with_lockers(4, 0, 2, lockers).await;
+        make_bucket_on_all(&disk_stores, bucket).await;
+        let (upload_id, parts) = stage_upload_with_create_opts(&set_disks, bucket, object, &[0x48; 4096], &create_opts).await;
+        let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        signaling.set_target(rustfs_lock::ObjectKey::new(RUSTFS_META_MULTIPART_BUCKET, upload_id_path.clone()));
+        let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;
+        let object_holder = set_disks
+            .new_ns_lock(bucket, object)
+            .await
+            .expect("object namespace lock should be created")
+            .get_write_lock(Duration::from_secs(5))
+            .await
+            .expect("test should hold the object lock");
+        let holder = set_disks
+            .new_ns_lock(RUSTFS_META_MULTIPART_BUCKET, &upload_id_path)
+            .await
+            .expect("upload namespace lock should be created")
+            .get_write_lock(Duration::from_secs(5))
+            .await
+            .expect("test should hold the upload lock");
+        signaling.wait_for_attempts(1).await;
+
+        let abort_store = set_disks.clone();
+        let abort_upload_id = upload_id.clone();
+        let abort = tokio::spawn(async move {
+            abort_store
+                .abort_multipart_upload(bucket, object, &abort_upload_id, &ObjectOptions::default())
+                .await
+        });
+        signaling.wait_for_attempts(2).await;
+
+        let complete_store = set_disks.clone();
+        let complete_upload_id = upload_id.clone();
+        let complete = tokio::spawn(async move {
+            complete_store
+                .complete_multipart_upload(bucket, object, &complete_upload_id, parts, &ObjectOptions::default())
+                .await
+        });
+        drop(holder);
+
+        abort
+            .await
+            .expect("abort task should not panic")
+            .expect("abort should win the upload finalization");
+        drop(object_holder);
+        let complete_err = complete
+            .await
+            .expect("completion task should not panic")
+            .expect_err("completion must observe the aborted upload");
+        assert!(matches!(complete_err, StorageError::InvalidUploadID(..)));
+        let object_err = set_disks
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect_err("abort-first must not publish an object");
+        assert!(matches!(object_err, StorageError::ObjectNotFound(..)));
+        assert!(matches!(
+            set_disks.check_upload_id_exists(bucket, object, &upload_id, false).await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+    }
+
     async fn assert_quorum_minus_one_retry_preserves_completable_part(
         disk_count: usize,
         parity: usize,
@@ -3039,6 +3192,81 @@ mod tests {
         .await;
     }
 
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn abort_and_complete_linearize_for_plain_sse_and_legacy_layouts() {
+        assert_complete_first_linearizes("multipart-complete-first-plain", "object", ObjectOptions::default()).await;
+        assert_abort_first_linearizes("multipart-abort-first-plain", "object", ObjectOptions::default()).await;
+
+        let encrypted_opts = ObjectOptions {
+            user_defined: HashMap::from([(SSEC_ALGORITHM_HEADER.to_string(), "AES256".to_string())]),
+            ..Default::default()
+        };
+        temp_env::async_with_vars([(crate::object_api::ENV_RUSTFS_ENCRYPTED_RANGE_SEEK, Some("true"))], async {
+            assert_complete_first_linearizes("multipart-complete-first-sse", "object", encrypted_opts.clone()).await;
+            assert_abort_first_linearizes("multipart-abort-first-sse", "object", encrypted_opts.clone()).await;
+        })
+        .await;
+        temp_env::async_with_vars([(crate::object_api::ENV_RUSTFS_ENCRYPTED_RANGE_SEEK, Some("false"))], async {
+            assert_complete_first_linearizes("multipart-complete-first-legacy", "object", encrypted_opts.clone()).await;
+            assert_abort_first_linearizes("multipart-abort-first-legacy", "object", encrypted_opts).await;
+        })
+        .await;
+    }
+
+    #[tokio::test]
+    async fn abort_enforces_delete_write_quorum_boundary() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-abort-delete-quorum";
+        let object = "object";
+        make_bucket_on_all(&disk_stores, bucket).await;
+        let quorum_upload = set_disks
+            .new_multipart_upload(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("multipart upload should be created");
+
+        let saved_disks = {
+            let mut disks = set_disks.disks.write().await;
+            let saved = disks.clone();
+            disks[3] = None;
+            saved
+        };
+        set_disks
+            .abort_multipart_upload(bucket, object, &quorum_upload.upload_id, &ObjectOptions::default())
+            .await
+            .expect("abort should succeed at the exact delete write quorum");
+        *set_disks.disks.write().await = saved_disks;
+        assert!(matches!(
+            set_disks
+                .check_upload_id_exists(bucket, object, &quorum_upload.upload_id, false)
+                .await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+
+        let below_quorum_upload = set_disks
+            .new_multipart_upload(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("second multipart upload should be created");
+        let saved_disks = {
+            let mut disks = set_disks.disks.write().await;
+            let saved = disks.clone();
+            disks[2] = None;
+            disks[3] = None;
+            saved
+        };
+        let err = set_disks
+            .abort_multipart_upload(bucket, object, &below_quorum_upload.upload_id, &ObjectOptions::default())
+            .await
+            .expect_err("abort must report a delete below write quorum");
+        assert!(matches!(err, StorageError::ErasureWriteQuorum));
+
+        *set_disks.disks.write().await = saved_disks;
+        set_disks
+            .check_upload_id_exists(bucket, object, &below_quorum_upload.upload_id, false)
+            .await
+            .expect("failed abort must leave quorum-visible staging on the restored disks");
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     #[serial]
     async fn complete_revalidates_layout_candidate_after_upload_lock() {
@@ -3170,6 +3398,17 @@ mod tests {
             tokio::task::yield_now().await;
             assert!(!abort.is_finished(), "abort must wait until completion releases the upload lock");
 
+            let list_store = set_disks.clone();
+            let list_upload_id = upload_id.clone();
+            let list = tokio::spawn(async move {
+                list_store
+                    .list_object_parts(bucket, object, &list_upload_id, None, MAX_PARTS_COUNT, &ObjectOptions::default())
+                    .await
+            });
+            signaling.wait_for_attempts(3).await;
+            tokio::task::yield_now().await;
+            assert!(!list.is_finished(), "ListParts must wait until completion releases the upload lock");
+
             barrier.release();
             complete
                 .await
@@ -3180,10 +3419,68 @@ mod tests {
                 .expect("abort task should not panic")
                 .expect_err("the committed upload should no longer exist when abort acquires the lock");
             assert!(matches!(abort_err, StorageError::InvalidUploadID(..)));
+            let list_err = list
+                .await
+                .expect("ListParts task should not panic")
+                .expect_err("the committed upload should no longer exist when ListParts acquires the lock");
+            assert!(matches!(list_err, StorageError::InvalidUploadID(..)));
         })
         .await;
     }
 
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn complete_validates_parts_after_an_inflight_upload_part_commit() {
+        let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+        let signaling = Arc::new(SignalingLockClient::new(Arc::new(LocalClient::with_manager(manager))));
+        let lockers: Vec<Arc<dyn LockClient>> = vec![signaling.clone()];
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks_with_lockers(4, 0, 2, lockers).await;
+        let bucket = "multipart-complete-put-part-race-bucket";
+        let object = "object";
+        make_bucket_on_all(&disk_stores, bucket).await;
+        let (upload_id, original_parts) =
+            stage_upload_with_create_opts(&set_disks, bucket, object, &[0x49; 4096], &ObjectOptions::default()).await;
+        let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        signaling.set_target(rustfs_lock::ObjectKey::new(RUSTFS_META_MULTIPART_BUCKET, upload_id_path));
+        let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;
+        let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::PutPartBeforeLockLost);
+
+        let put_store = set_disks.clone();
+        let put_upload_id = upload_id.clone();
+        let put = tokio::spawn(async move {
+            let mut reader = PutObjReader::from_vec(vec![0x4a; 4096]);
+            put_store
+                .put_object_part(bucket, object, &put_upload_id, 1, &mut reader, &ObjectOptions::default())
+                .await
+        });
+        barrier.wait_until_paused().await;
+
+        let complete_store = set_disks.clone();
+        let complete_upload_id = upload_id.clone();
+        let complete = tokio::spawn(async move {
+            complete_store
+                .complete_multipart_upload(bucket, object, &complete_upload_id, original_parts, &ObjectOptions::default())
+                .await
+        });
+        signaling.wait_for_attempts(2).await;
+        tokio::task::yield_now().await;
+        assert!(!complete.is_finished(), "completion must wait for the UploadPart commit lock");
+
+        barrier.release();
+        put.await
+            .expect("UploadPart task should not panic")
+            .expect("UploadPart replacement should commit");
+        let err = complete
+            .await
+            .expect("completion task should not panic")
+            .expect_err("completion must reject the stale ETag after UploadPart wins");
+        assert!(matches!(err, StorageError::InvalidPart(..)));
+        set_disks
+            .list_object_parts(bucket, object, &upload_id, None, MAX_PARTS_COUNT, &ObjectOptions::default())
+            .await
+            .expect("failed completion must leave the upload retryable");
+    }
+
     #[tokio::test(start_paused = true)]
     #[serial]
     async fn complete_fences_upload_lock_loss_before_commit() {

crates/ecstore/src/store/object.rs

@@ -2601,10 +2601,10 @@ mod tests {
     // (backlog#1304): restore entry no longer serializes on the object lock.
     // The replacement semantics — non-blocking reads during the copy-back and
     // fast rejection of a concurrent restore — are covered end-to-end by
-    // `restore_object_usecase_reports_ongoing_conflict_and_completion`
-    // (rustfs/src/app/lifecycle_transition_api_test.rs) and at the lock level
-    // by the accept-guard test below; restore-vs-reader data protection lives
-    // in the inner put_object/complete_multipart_upload commit locks.
+    // `restore_object_usecase_reports_ongoing_conflict`
+    // (rustfs/src/app/lifecycle_transition_api_test.rs), while the SetDisks
+    // transition matrix covers the final local commit. Restore-vs-reader data
+    // protection lives in the inner put_object/complete_multipart_upload locks.
     #[tokio::test]
     #[serial_test::serial]
     async fn restore_accept_guard_serializes_concurrent_accepts() {

crates/protocols/Cargo.toml

@@ -134,6 +134,7 @@ socket2 = { workspace = true, optional = true, features = ["all"] }
 [dev-dependencies]
 tempfile = { workspace = true }
 proptest = "1"
+rustfs-test-utils = { workspace = true }
 tracing-subscriber = { workspace = true, features = ["env-filter", "time"] }
 tokio = { workspace = true, features = ["test-util", "macros", "fs"] }
 

crates/protocols/src/swift/account.rs

@@ -16,12 +16,11 @@
 
 use super::storage_api::account::{BucketOperations, MakeBucketOptions};
 use super::{SwiftError, SwiftResult};
-use super::{get_swift_bucket_metadata, resolve_swift_object_store_handle, set_swift_bucket_metadata};
+use super::{get_swift_bucket_metadata, resolve_swift_object_store_handle, update_swift_bucket_tagging, validate_metadata};
 use rustfs_credentials::Credentials;
 use s3s::dto::{Tag, Tagging};
 use sha2::{Digest, Sha256};
 use std::collections::HashMap;
-use time;
 
 /// Validate that the authenticated user has access to the requested account
 ///
@@ -148,15 +147,33 @@ pub async fn get_account_metadata(account: &str, _credentials: &Option<Credentia
 /// Updates account-level metadata such as TempURL keys.
 /// Only updates swift-account-meta-* tags, preserving other tags.
 ///
+/// The caller must own the account: this metadata holds the account's TempURL
+/// signing key, so writing it for someone else's account would let the writer
+/// mint valid pre-signed URLs against that account's objects. Reads
+/// ([`get_account_metadata`]) stay unauthenticated on purpose — TempURL
+/// signature validation has to resolve the key before any credentials exist.
+///
 /// # Arguments
 /// * `account` - Account identifier
 /// * `metadata` - Metadata key-value pairs to store (keys will be prefixed with `swift-account-meta-`)
-/// * `credentials` - S3 credentials
+/// * `credentials` - Keystone credentials of the caller
 pub async fn update_account_metadata(
     account: &str,
     metadata: &HashMap<String, String>,
-    _credentials: &Option<Credentials>,
+    credentials: &Option<Credentials>,
 ) -> SwiftResult<()> {
+    let Some(credentials) = credentials.as_ref() else {
+        return Err(SwiftError::Unauthorized(
+            "Keystone authentication required to update account metadata".to_string(),
+        ));
+    };
+    validate_account_access(account, credentials)?;
+
+    // These tags are persisted into the bucket metadata file, which every
+    // later config write rewrites in full — so unbounded metadata inflates
+    // the cost of unrelated writes for the life of the account.
+    validate_metadata(metadata)?;
+
     let bucket_name = get_account_metadata_bucket_name(account);
 
     let Some(store) = resolve_swift_object_store_handle() else {
@@ -173,57 +190,30 @@ pub async fn update_account_metadata(
             .map_err(|e| SwiftError::InternalServerError(format!("Failed to create account metadata bucket: {}", e)))?;
     }
 
-    // Load current bucket metadata
-    let bucket_meta = get_swift_bucket_metadata(&bucket_name)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to load bucket metadata: {}", e)))?;
+    // Rewrite the persisted tags: replace swift-account-meta-* tags with the
+    // new metadata while preserving other tags. An empty result clears the
+    // tagging config.
+    update_swift_bucket_tagging(bucket_name, |current| {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
 
-    let mut bucket_meta_clone = (*bucket_meta).clone();
-
-    // Get existing tags, preserving non-Swift tags
-    let mut existing_tagging = bucket_meta_clone
-        .tagging_config
-        .clone()
-        .unwrap_or_else(|| Tagging { tag_set: vec![] });
-
-    // Remove old swift-account-meta-* tags while preserving other tags
-    existing_tagging.tag_set.retain(|tag| {
-        if let Some(key) = &tag.key {
-            !key.starts_with("swift-account-meta-")
-        } else {
-            true
-        }
-    });
-
-    // Add new metadata tags
-    for (key, value) in metadata {
-        existing_tagging.tag_set.push(Tag {
-            key: Some(format!("swift-account-meta-{}", key)),
-            value: Some(value.clone()),
+        tagging.tag_set.retain(|tag| {
+            if let Some(key) = &tag.key {
+                !key.starts_with("swift-account-meta-")
+            } else {
+                true
+            }
         });
-    }
 
-    let now = time::OffsetDateTime::now_utc();
+        for (key, value) in metadata {
+            tagging.tag_set.push(Tag {
+                key: Some(format!("swift-account-meta-{}", key)),
+                value: Some(value.clone()),
+            });
+        }
 
-    if existing_tagging.tag_set.is_empty() {
-        // No tags remain; clear tagging config
-        bucket_meta_clone.tagging_config_xml = Vec::new();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = None;
-    } else {
-        // Serialize tags to XML
-        let tagging_xml = quick_xml::se::to_string(&existing_tagging)
-            .map_err(|e| SwiftError::InternalServerError(format!("Failed to serialize tags: {}", e)))?;
-
-        bucket_meta_clone.tagging_config_xml = tagging_xml.into_bytes();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = Some(existing_tagging);
-    }
-
-    // Save updated metadata
-    set_swift_bucket_metadata(bucket_name.clone(), bucket_meta_clone)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to save metadata: {}", e)))?;
+        tagging
+    })
+    .await?;
 
     Ok(())
 }

crates/protocols/src/swift/container.rs

@@ -22,7 +22,10 @@ use super::storage_api::container::{
 };
 use super::types::Container;
 use super::{SwiftError, SwiftResult};
-use super::{get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, set_swift_bucket_metadata};
+use super::{
+    get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, update_swift_bucket_tagging,
+    validate_metadata,
+};
 use rustfs_credentials::Credentials;
 use s3s::dto::{Tag, Tagging};
 use sha2::{Digest, Sha256};
@@ -483,6 +486,11 @@ pub async fn update_container_metadata(
     // Validate container name
     validate_container_name(container)?;
 
+    // These tags are persisted into the bucket metadata file, which every
+    // later config write rewrites in full — so unbounded metadata inflates
+    // the cost of unrelated writes for the life of the container.
+    validate_metadata(&metadata)?;
+
     // Create mapper with default config (tenant prefixing enabled)
     let mapper = ContainerMapper::default();
 
@@ -506,57 +514,30 @@ pub async fn update_container_metadata(
             }
         })?;
 
-    // Load current bucket metadata
-    let bucket_meta = get_swift_bucket_metadata(&bucket_name)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to load bucket metadata: {}", e)))?;
+    // Rewrite the persisted tags: replace swift-meta-* tags with the new
+    // metadata while preserving non-Swift tags. An empty result clears the
+    // tagging config.
+    update_swift_bucket_tagging(bucket_name, |current| {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
 
-    let mut bucket_meta_clone = (*bucket_meta).clone();
+        tagging.tag_set.retain(|tag| {
+            if let Some(key) = &tag.key {
+                !key.starts_with("swift-meta-")
+            } else {
+                true // Keep tags with no key (shouldn't happen, but be safe)
+            }
+        });
 
-    // Get existing tags, preserving non-Swift tags
-    let mut existing_tagging = bucket_meta_clone
-        .tagging_config
-        .clone()
-        .unwrap_or_else(|| Tagging { tag_set: vec![] });
-
-    // Remove old swift-meta-* tags while preserving other tags
-    existing_tagging.tag_set.retain(|tag| {
-        if let Some(key) = &tag.key {
-            !key.starts_with("swift-meta-")
-        } else {
-            true // Keep tags with no key (shouldn't happen, but be safe)
+        if let Some(mut new_tagging) = swift_metadata_to_s3_tags(&metadata) {
+            tagging.tag_set.append(&mut new_tagging.tag_set);
         }
-    });
+        // If metadata.is_empty() and swift_metadata_to_s3_tags returns None,
+        // we've already removed swift-meta-* tags above, so only non-Swift
+        // tags remain
 
-    // Add new Swift metadata tags if provided
-    if let Some(mut new_tagging) = swift_metadata_to_s3_tags(&metadata) {
-        // Merge: existing non-Swift tags + new Swift tags
-        existing_tagging.tag_set.append(&mut new_tagging.tag_set);
-    }
-    // If metadata.is_empty() and swift_metadata_to_s3_tags returns None,
-    // we've already removed swift-meta-* tags above, so only non-Swift tags remain
-
-    let now = time::OffsetDateTime::now_utc();
-
-    if existing_tagging.tag_set.is_empty() {
-        // No tags remain after removing swift-meta-* tags; clear tagging config
-        bucket_meta_clone.tagging_config_xml = Vec::new();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = None;
-    } else {
-        // Serialize the merged tags to XML
-        let tagging_xml = quick_xml::se::to_string(&existing_tagging)
-            .map_err(|e| SwiftError::InternalServerError(format!("Failed to serialize tags: {}", e)))?;
-
-        bucket_meta_clone.tagging_config_xml = tagging_xml.into_bytes();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = Some(existing_tagging);
-    }
-
-    // Save updated metadata
-    set_swift_bucket_metadata(bucket_name, bucket_meta_clone)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to save metadata: {}", e)))?;
+        tagging
+    })
+    .await?;
 
     Ok(())
 }
@@ -819,44 +800,23 @@ pub async fn enable_versioning(
             }
         })?;
 
-    // Load current bucket metadata
-    let bucket_meta = get_swift_bucket_metadata(&bucket_name)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to load bucket metadata: {}", e)))?;
+    // Rewrite the persisted tags: replace any versioning tag with the new
+    // archive location while preserving all other tags.
+    update_swift_bucket_tagging(bucket_name, |current| {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
 
-    let mut bucket_meta_clone = (*bucket_meta).clone();
+        tagging
+            .tag_set
+            .retain(|tag| tag.key.as_deref() != Some("swift-versions-location"));
 
-    // Get existing tags
-    let mut existing_tagging = bucket_meta_clone
-        .tagging_config
-        .clone()
-        .unwrap_or_else(|| Tagging { tag_set: vec![] });
+        tagging.tag_set.push(Tag {
+            key: Some("swift-versions-location".to_string()),
+            value: Some(archive_container.to_string()), // Store Swift container name, not S3 bucket name
+        });
 
-    // Remove old versioning tag if present
-    existing_tagging
-        .tag_set
-        .retain(|tag| tag.key.as_deref() != Some("swift-versions-location"));
-
-    // Add new versioning tag
-    existing_tagging.tag_set.push(Tag {
-        key: Some("swift-versions-location".to_string()),
-        value: Some(archive_container.to_string()), // Store Swift container name, not S3 bucket name
-    });
-
-    let now = time::OffsetDateTime::now_utc();
-
-    // Serialize tags to XML
-    let tagging_xml = quick_xml::se::to_string(&existing_tagging)
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to serialize tags: {}", e)))?;
-
-    bucket_meta_clone.tagging_config_xml = tagging_xml.into_bytes();
-    bucket_meta_clone.tagging_config_updated_at = now;
-    bucket_meta_clone.tagging_config = Some(existing_tagging);
-
-    // Save updated metadata
-    set_swift_bucket_metadata(bucket_name, bucket_meta_clone)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to save metadata: {}", e)))?;
+        tagging
+    })
+    .await?;
 
     Ok(())
 }
@@ -882,50 +842,38 @@ pub async fn disable_versioning(account: &str, container: &str, credentials: &Cr
     let mapper = ContainerMapper::default();
     let bucket_name = mapper.swift_to_s3_bucket(container, &project_id);
 
-    // Verify container exists
-    let Some(_store) = resolve_swift_object_store_handle() else {
+    let Some(store) = resolve_swift_object_store_handle() else {
         return Err(SwiftError::InternalServerError("Storage layer not initialized".to_string()));
     };
 
-    // Load current bucket metadata
-    let bucket_meta = get_swift_bucket_metadata(&bucket_name)
+    // Verify container exists. Without this the rewrite below would persist a
+    // fabricated default for a container that does not exist: the metadata
+    // loader turns "no metadata on disk" into a fresh BucketMetadata, and
+    // writing that creates an orphan .metadata.bin and caches a fabricated
+    // default as authoritative.
+    store
+        .get_bucket_info(&bucket_name, &BucketOptions::default())
         .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to load bucket metadata: {}", e)))?;
+        .map_err(|e| {
+            if e.to_string().contains("not found") || e.to_string().contains("NoSuchBucket") {
+                SwiftError::NotFound(format!("Container '{}' not found", container))
+            } else {
+                sanitize_storage_error("Container verification", e)
+            }
+        })?;
 
-    let mut bucket_meta_clone = (*bucket_meta).clone();
+    // Rewrite the persisted tags: drop the versioning tag while preserving
+    // all other tags. An empty result clears the tagging config.
+    update_swift_bucket_tagging(bucket_name, |current| {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
 
-    // Get existing tags
-    let mut existing_tagging = bucket_meta_clone
-        .tagging_config
-        .clone()
-        .unwrap_or_else(|| Tagging { tag_set: vec![] });
+        tagging
+            .tag_set
+            .retain(|tag| tag.key.as_deref() != Some("swift-versions-location"));
 
-    // Remove versioning tag
-    existing_tagging
-        .tag_set
-        .retain(|tag| tag.key.as_deref() != Some("swift-versions-location"));
-
-    let now = time::OffsetDateTime::now_utc();
-
-    if existing_tagging.tag_set.is_empty() {
-        // No tags remain; clear tagging config
-        bucket_meta_clone.tagging_config_xml = Vec::new();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = None;
-    } else {
-        // Serialize remaining tags to XML
-        let tagging_xml = quick_xml::se::to_string(&existing_tagging)
-            .map_err(|e| SwiftError::InternalServerError(format!("Failed to serialize tags: {}", e)))?;
-
-        bucket_meta_clone.tagging_config_xml = tagging_xml.into_bytes();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = Some(existing_tagging);
-    }
-
-    // Save updated metadata
-    set_swift_bucket_metadata(bucket_name, bucket_meta_clone)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to save metadata: {}", e)))?;
+        tagging
+    })
+    .await?;
 
     Ok(())
 }
@@ -1050,65 +998,37 @@ pub async fn set_container_acl(
             }
         })?;
 
-    // Load current bucket metadata
-    let bucket_meta = get_swift_bucket_metadata(&bucket_name)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to load bucket metadata: {}", e)))?;
+    // Rewrite the persisted tags: replace the ACL tags with the new grants
+    // while preserving all other tags. An empty result clears the tagging
+    // config.
+    update_swift_bucket_tagging(bucket_name, |current| {
+        let mut tagging = current.cloned().unwrap_or_else(|| Tagging { tag_set: vec![] });
 
-    let mut bucket_meta_clone = (*bucket_meta).clone();
+        tagging
+            .tag_set
+            .retain(|tag| tag.key.as_deref() != Some("swift-acl-read") && tag.key.as_deref() != Some("swift-acl-write"));
 
-    // Get existing tags
-    let mut existing_tagging = bucket_meta_clone
-        .tagging_config
-        .clone()
-        .unwrap_or_else(|| Tagging { tag_set: vec![] });
+        if let Some(read) = read_acl
+            && !read.trim().is_empty()
+        {
+            tagging.tag_set.push(Tag {
+                key: Some("swift-acl-read".to_string()),
+                value: Some(read.to_string()),
+            });
+        }
 
-    // Remove old ACL tags
-    existing_tagging
-        .tag_set
-        .retain(|tag| tag.key.as_deref() != Some("swift-acl-read") && tag.key.as_deref() != Some("swift-acl-write"));
+        if let Some(write) = write_acl
+            && !write.trim().is_empty()
+        {
+            tagging.tag_set.push(Tag {
+                key: Some("swift-acl-write".to_string()),
+                value: Some(write.to_string()),
+            });
+        }
 
-    // Add new read ACL tag if provided
-    if let Some(read) = read_acl
-        && !read.trim().is_empty()
-    {
-        existing_tagging.tag_set.push(Tag {
-            key: Some("swift-acl-read".to_string()),
-            value: Some(read.to_string()),
-        });
-    }
-
-    // Add new write ACL tag if provided
-    if let Some(write) = write_acl
-        && !write.trim().is_empty()
-    {
-        existing_tagging.tag_set.push(Tag {
-            key: Some("swift-acl-write".to_string()),
-            value: Some(write.to_string()),
-        });
-    }
-
-    let now = time::OffsetDateTime::now_utc();
-
-    if existing_tagging.tag_set.is_empty() {
-        // No tags remain; clear tagging config
-        bucket_meta_clone.tagging_config_xml = Vec::new();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = None;
-    } else {
-        // Serialize tags to XML
-        let tagging_xml = quick_xml::se::to_string(&existing_tagging)
-            .map_err(|e| SwiftError::InternalServerError(format!("Failed to serialize tags: {}", e)))?;
-
-        bucket_meta_clone.tagging_config_xml = tagging_xml.into_bytes();
-        bucket_meta_clone.tagging_config_updated_at = now;
-        bucket_meta_clone.tagging_config = Some(existing_tagging);
-    }
-
-    // Save updated metadata
-    set_swift_bucket_metadata(bucket_name, bucket_meta_clone)
-        .await
-        .map_err(|e| SwiftError::InternalServerError(format!("Failed to save metadata: {}", e)))?;
+        tagging
+    })
+    .await?;
 
     debug!(
         "Set ACLs for container {}/{}: read={:?}, write={:?}",

crates/protocols/src/swift/mod.rs

@@ -58,10 +58,53 @@ pub mod versioning;
 
 pub use errors::{SwiftError, SwiftResult};
 pub use router::{SwiftRoute, SwiftRouter};
+
+/// Maximum number of metadata headers allowed per resource (Swift standard)
+pub(crate) const MAX_METADATA_COUNT: usize = 90;
+
+/// Maximum size in bytes for a single metadata value (Swift standard)
+pub(crate) const MAX_METADATA_VALUE_SIZE: usize = 256;
+
+/// Validate metadata against Swift limits
+///
+/// Checks that:
+/// - Total number of metadata entries doesn't exceed MAX_METADATA_COUNT
+/// - Individual metadata values don't exceed MAX_METADATA_VALUE_SIZE
+///
+/// Applies to object, container and account metadata alike: all three are
+/// persisted, and container/account metadata additionally lands in the
+/// bucket metadata file that every later config write rewrites in full.
+///
+/// Returns error if limits are exceeded.
+pub(crate) fn validate_metadata(metadata: &std::collections::HashMap<String, String>) -> SwiftResult<()> {
+    // Check total metadata count
+    if metadata.len() > MAX_METADATA_COUNT {
+        return Err(SwiftError::BadRequest(format!(
+            "Too many metadata headers: {} (max: {})",
+            metadata.len(),
+            MAX_METADATA_COUNT
+        )));
+    }
+
+    // Check individual value sizes
+    for (key, value) in metadata.iter() {
+        if value.len() > MAX_METADATA_VALUE_SIZE {
+            return Err(SwiftError::BadRequest(format!(
+                "Metadata value for '{}' too large: {} bytes (max: {} bytes)",
+                key,
+                value.len(),
+                MAX_METADATA_VALUE_SIZE
+            )));
+        }
+    }
+
+    Ok(())
+}
+
 // Note: Container, Object, and SwiftMetadata types used by Swift implementation
 pub use storage_api::public_api::{SwiftGetObjectReader, SwiftObjectInfo, SwiftObjectOptions, SwiftPutObjReader};
 pub(crate) use storage_api::public_api::{
-    get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, set_swift_bucket_metadata,
+    get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, update_swift_bucket_tagging,
 };
 #[allow(unused_imports)]
 pub use types::{Container, Object, SwiftMetadata};

crates/protocols/src/swift/object.rs

@@ -53,7 +53,7 @@ use super::account::validate_account_access;
 use super::container::ContainerMapper;
 use super::expiration_worker::{track_object_expiration, untrack_object_expiration};
 use super::storage_api::object::{BucketOperations, BucketOptions, HTTPRangeSpec, ObjectIO as _, ObjectOperations as _};
-use super::{SwiftError, SwiftResult, resolve_swift_object_store_handle};
+use super::{SwiftError, SwiftResult, resolve_swift_object_store_handle, validate_metadata};
 use axum::http::HeaderMap;
 use rustfs_credentials::Credentials;
 use rustfs_rio::HashReader;
@@ -68,12 +68,6 @@ const LOG_SUBSYSTEM_SWIFT_OBJECT: &str = "swift_object";
 const EVENT_SWIFT_OBJECT_STORAGE_STATE: &str = "swift_object_storage_state";
 const SWIFT_DELETE_AT_METADATA: &str = "x-delete-at";
 
-/// Maximum number of metadata headers allowed per object (Swift standard)
-const MAX_METADATA_COUNT: usize = 90;
-
-/// Maximum size in bytes for a single metadata value (Swift standard)
-const MAX_METADATA_VALUE_SIZE: usize = 256;
-
 /// Maximum object size in bytes (5GB - Swift default)
 const MAX_OBJECT_SIZE: i64 = 5 * 1024 * 1024 * 1024;
 
@@ -234,38 +228,6 @@ impl Default for ObjectKeyMapper {
     }
 }
 
-/// Validate metadata against Swift limits
-///
-/// Checks that:
-/// - Total number of metadata entries doesn't exceed MAX_METADATA_COUNT
-/// - Individual metadata values don't exceed MAX_METADATA_VALUE_SIZE
-///
-/// Returns error if limits are exceeded.
-fn validate_metadata(metadata: &HashMap<String, String>) -> SwiftResult<()> {
-    // Check total metadata count
-    if metadata.len() > MAX_METADATA_COUNT {
-        return Err(SwiftError::BadRequest(format!(
-            "Too many metadata headers: {} (max: {})",
-            metadata.len(),
-            MAX_METADATA_COUNT
-        )));
-    }
-
-    // Check individual value sizes
-    for (key, value) in metadata.iter() {
-        if value.len() > MAX_METADATA_VALUE_SIZE {
-            return Err(SwiftError::BadRequest(format!(
-                "Metadata value for '{}' too large: {} bytes (max: {} bytes)",
-                key,
-                value.len(),
-                MAX_METADATA_VALUE_SIZE
-            )));
-        }
-    }
-
-    Ok(())
-}
-
 fn metadata_delete_at(metadata: &HashMap<String, String>) -> Option<u64> {
     metadata
         .get(SWIFT_DELETE_AT_METADATA)

crates/protocols/src/swift/storage_api.rs

@@ -15,14 +15,19 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
+use rustfs_ecstore::api::bucket::metadata::BUCKET_TAGGING_CONFIG;
 pub(crate) use rustfs_ecstore::api::bucket::metadata::BucketMetadata as SwiftBucketMetadata;
-use rustfs_ecstore::api::bucket::metadata_sys::{
-    get as get_swift_bucket_metadata_from_backend, set_bucket_metadata as set_swift_bucket_metadata_in_backend,
-};
+use rustfs_ecstore::api::bucket::metadata_sys::{get as get_swift_bucket_metadata_from_backend, update_config_with};
+use rustfs_ecstore::api::bucket::utils::serialize as serialize_bucket_config;
+use rustfs_ecstore::api::error::Error as SwiftStorageError;
 pub(crate) use rustfs_ecstore::api::error::Result as SwiftStorageResult;
+use rustfs_ecstore::api::notification::get_global_notification_sys;
 pub(crate) use rustfs_ecstore::api::runtime::object_store_handle as resolve_swift_object_store_handle;
 use rustfs_ecstore::api::storage::ECStore as SwiftStore;
 use rustfs_storage_api as storage_contracts;
+use s3s::dto::Tagging;
+
+use super::{SwiftError, SwiftResult};
 
 pub(crate) mod account {
     pub(crate) use super::storage_contracts::{BucketOperations, MakeBucketOptions};
@@ -45,7 +50,7 @@ pub(crate) mod object {
 pub(crate) mod public_api {
     pub use super::{SwiftGetObjectReader, SwiftObjectInfo, SwiftObjectOptions, SwiftPutObjReader};
     pub(crate) use super::{
-        get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, set_swift_bucket_metadata,
+        get_swift_bucket_metadata, get_swift_bucket_usage, resolve_swift_object_store_handle, update_swift_bucket_tagging,
     };
 }
 
@@ -53,6 +58,15 @@ pub(crate) mod versioning {
     pub(crate) use super::storage_contracts::{ListOperations, ObjectOperations};
 }
 
+const LOG_COMPONENT_PROTOCOLS: &str = "protocols";
+const LOG_SUBSYSTEM_SWIFT_STORAGE: &str = "swift_storage";
+const EVENT_SWIFT_BUCKET_TAGGING_UPDATE: &str = "swift_bucket_tagging_update";
+
+/// Marks the refusal to rewrite an unreadable persisted tagging config, so the
+/// caller can turn it into an actionable client error rather than a generic
+/// storage failure. Carried through the ecstore error, which is a string type.
+const UNREADABLE_TAGGING_SENTINEL: &str = "swift: persisted tagging config could not be parsed";
+
 pub type SwiftGetObjectReader = <SwiftStore as storage_contracts::ObjectIO>::GetObjectReader;
 pub type SwiftObjectInfo = <SwiftStore as storage_contracts::ObjectOperations>::ObjectInfo;
 pub type SwiftObjectOptions = <SwiftStore as storage_contracts::ObjectOperations>::ObjectOptions;
@@ -62,8 +76,80 @@ pub(crate) async fn get_swift_bucket_metadata(bucket: &str) -> SwiftStorageResul
     get_swift_bucket_metadata_from_backend(bucket).await
 }
 
-pub(crate) async fn set_swift_bucket_metadata(bucket: String, metadata: SwiftBucketMetadata) -> SwiftStorageResult<()> {
-    set_swift_bucket_metadata_in_backend(bucket, metadata).await
+/// Rewrite the bucket's tagging config through the persisting
+/// bucket-metadata path.
+///
+/// `rewrite` sees the tag set currently persisted on disk (`None` when the
+/// bucket has none) and returns the full replacement; an empty tag set
+/// clears the config. The read-modify-write runs under the bucket metadata
+/// system's write guard — serialized against every other config update —
+/// and the result is written to the bucket metadata file before the cache
+/// is refreshed, so a Swift metadata POST survives process restarts and
+/// disk-truth reloads. Peers are then told to reload, matching what the S3
+/// handlers do after a config write.
+///
+/// Storage failures are logged in full and reported to the client as a
+/// generic error: these now carry real disk and quorum detail, which does not
+/// belong in a Swift response body. The one exception is an unreadable
+/// persisted config, which is reported specifically because the operator has
+/// to act on it.
+pub(crate) async fn update_swift_bucket_tagging<F>(bucket: String, rewrite: F) -> SwiftResult<()>
+where
+    F: FnOnce(Option<&Tagging>) -> Tagging + Send,
+{
+    let result = update_config_with(&bucket, BUCKET_TAGGING_CONFIG, |bm| {
+        // Merging onto an unparseable tag set would silently drop every tag
+        // the bucket has — including the container ACL and versioning tags —
+        // because the rewrite closures treat "no parsed tags" as "no tags".
+        // Refuse instead: the persisted config is intact, just unreadable.
+        if !bm.tagging_config_xml.is_empty() && bm.tagging_config.is_none() {
+            return Err(SwiftStorageError::other(UNREADABLE_TAGGING_SENTINEL));
+        }
+
+        let tagging = rewrite(bm.tagging_config.as_ref());
+        if tagging.tag_set.is_empty() {
+            Ok(Vec::new())
+        } else {
+            // The S3 XML serializer, not quick_xml: the metadata loader's
+            // parse step must be able to round-trip what we persist.
+            serialize_bucket_config(&tagging)
+                .map_err(|e| SwiftStorageError::other(format!("failed to serialize bucket tagging: {e}")))
+        }
+    })
+    .await;
+
+    if let Err(err) = result {
+        let unreadable = err.to_string().contains(UNREADABLE_TAGGING_SENTINEL);
+        tracing::error!(
+            event = EVENT_SWIFT_BUCKET_TAGGING_UPDATE,
+            component = LOG_COMPONENT_PROTOCOLS,
+            subsystem = LOG_SUBSYSTEM_SWIFT_STORAGE,
+            bucket = %bucket,
+            error = %err,
+            reason = if unreadable { "unreadable_persisted_config" } else { "storage_failure" },
+            result = "failed",
+            "swift bucket tagging update failed"
+        );
+        // A Swift-only client has no way to repair this itself, so say what
+        // happened and name the remedy instead of a bare storage error.
+        return Err(if unreadable {
+            SwiftError::Conflict(format!(
+                "The persisted tagging configuration for container store '{bucket}' cannot be parsed, so metadata cannot be updated without discarding it. Reset it with the S3 DeleteBucketTagging API."
+            ))
+        } else {
+            SwiftError::InternalServerError("Metadata update operation failed".to_string())
+        });
+    }
+
+    if let Some(notification_sys) = get_global_notification_sys() {
+        tokio::spawn(async move {
+            if let Err(err) = notification_sys.load_bucket_metadata(&bucket).await {
+                tracing::warn!(bucket = %bucket, error = %err, "failed to notify peers after swift bucket tagging update");
+            }
+        });
+    }
+
+    Ok(())
 }
 
 pub(crate) async fn get_swift_bucket_usage() -> SwiftStorageResult<Option<HashMap<String, (u64, u64)>>> {

crates/protocols/tests/ecstore_test_compat/mod.rs

@@ -0,0 +1,25 @@
+// 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.
+
+//! ECStore facade boundary for the protocols integration tests.
+//!
+//! The Swift tests need to drive bucket-metadata reloads the way the peer
+//! `LoadBucketMetadata` RPC does. Everything they touch from `rustfs_ecstore`
+//! is aliased here so the tests themselves hold no raw facade subpaths, the
+//! same boundary `crates/protocols/src/swift/storage_api.rs` provides for the
+//! Swift implementation.
+
+pub use rustfs_ecstore::api::bucket::metadata::load_bucket_metadata;
+pub use rustfs_ecstore::api::bucket::metadata_sys::{get as get_bucket_metadata, set_bucket_metadata};
+pub use rustfs_ecstore::api::runtime::object_store_handle;

crates/protocols/tests/swift_metadata_persistence.rs

@@ -0,0 +1,312 @@
+// 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.
+
+//! Regression tests: a Swift metadata POST must be persisted to the bucket
+//! metadata file, not just the in-memory cache. The metadata has to survive
+//! the disk-truth reloads performed by peer LoadBucketMetadata notifications
+//! and the periodic refresh loop — and, transitively, a process restart.
+
+#![cfg(feature = "swift")]
+
+use std::collections::HashMap;
+
+use rustfs_credentials::Credentials;
+use rustfs_protocols::swift::SwiftError;
+use rustfs_protocols::swift::container::{ContainerMapper, update_container_metadata};
+use rustfs_protocols::swift::{account, container};
+use rustfs_test_utils::TestECStoreEnv;
+use serde_json::json;
+use sha2::{Digest, Sha256};
+
+mod ecstore_test_compat;
+use ecstore_test_compat::{get_bucket_metadata, load_bucket_metadata, object_store_handle, set_bucket_metadata};
+
+fn keystone_credentials(project_id: &str) -> Credentials {
+    let mut claims = HashMap::new();
+    claims.insert("keystone_project_id".to_string(), json!(project_id));
+    claims.insert("keystone_roles".to_string(), json!(["member"]));
+
+    Credentials {
+        access_key: "keystone:swift-test".to_string(),
+        claims: Some(claims),
+        ..Default::default()
+    }
+}
+
+/// The account-metadata bucket name scheme from `swift::account`
+/// (`swift-account-{sha256(account)[0..16]}`), mirrored here so the test can
+/// reload that bucket's metadata from disk.
+fn account_metadata_bucket_name(account: &str) -> String {
+    let mut hasher = Sha256::new();
+    hasher.update(account.as_bytes());
+    let hash = hex::encode(hasher.finalize());
+    format!("swift-account-{}", &hash[0..16])
+}
+
+/// Replace the cached bucket metadata with what is actually on disk — the
+/// same thing a peer LoadBucketMetadata notification or the periodic refresh
+/// loop does. Before the fix this silently discarded every Swift metadata
+/// POST, because those writes only ever touched the cache.
+async fn reload_bucket_metadata_from_disk(bucket: &str) {
+    let store = object_store_handle().expect("test store should be published");
+    let bm = load_bucket_metadata(store, bucket)
+        .await
+        .expect("bucket metadata should load from disk");
+    set_bucket_metadata(bucket.to_string(), bm)
+        .await
+        .expect("reloaded metadata should install");
+}
+
+/// The `swift-meta-*` tags currently persisted for a container, keyed the way
+/// a Swift client sees them. `get_container_metadata` would be the natural
+/// reader, but it additionally requires a data-usage snapshot for the object
+/// count and byte total, which a bare test store has none of — and that is
+/// orthogonal to whether the metadata itself was persisted.
+async fn persisted_container_metadata(bucket: &str) -> HashMap<String, String> {
+    let bm = get_bucket_metadata(bucket).await.expect("bucket metadata should be cached");
+    let mut out = HashMap::new();
+    if let Some(tagging) = &bm.tagging_config {
+        for tag in &tagging.tag_set {
+            if let (Some(key), Some(value)) = (&tag.key, &tag.value)
+                && let Some(meta_key) = key.strip_prefix("swift-meta-")
+            {
+                out.insert(meta_key.to_string(), value.clone());
+            }
+        }
+    }
+    out
+}
+
+/// Every scenario that needs a store runs against ONE environment: the Swift
+/// handlers resolve the ambient object store and bucket-metadata system, so a
+/// second `TestECStoreEnv` in this process would race the first.
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn swift_metadata_writes_are_durable() {
+    let env = TestECStoreEnv::builder().prefix("swift_meta_persist").build().await;
+
+    posts_survive_disk_truth_reload(&env).await;
+    tag_writers_preserve_each_others_state(&env).await;
+    versioning_writes_reject_missing_containers().await;
+}
+
+async fn posts_survive_disk_truth_reload(env: &TestECStoreEnv) {
+    // --- Container metadata POST (X-Container-Meta-*) ---
+    let project_id = "swiftpersistproj";
+    let swift_account = format!("AUTH_{project_id}");
+    let credentials = keystone_credentials(project_id);
+    let swift_container = "photos";
+    let bucket = ContainerMapper::default().swift_to_s3_bucket(swift_container, project_id);
+    env.make_bucket(&bucket, false).await;
+
+    let mut metadata = HashMap::new();
+    metadata.insert("color".to_string(), "blue".to_string());
+    update_container_metadata(&swift_account, swift_container, &credentials, metadata)
+        .await
+        .expect("container metadata POST should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    let container_meta = persisted_container_metadata(&bucket).await;
+    assert_eq!(
+        container_meta.get("color").map(String::as_str),
+        Some("blue"),
+        "container metadata POST must survive a disk-truth metadata reload"
+    );
+
+    // A follow-up POST replaces the Swift metadata and that replacement must
+    // survive a reload too (the rewrite merges against disk state, so the
+    // previous value must actually be gone).
+    let mut metadata = HashMap::new();
+    metadata.insert("season".to_string(), "summer".to_string());
+    update_container_metadata(&swift_account, swift_container, &credentials, metadata)
+        .await
+        .expect("second container metadata POST should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    let container_meta = persisted_container_metadata(&bucket).await;
+    assert_eq!(container_meta.get("season").map(String::as_str), Some("summer"));
+    assert!(
+        !container_meta.contains_key("color"),
+        "replaced container metadata must not resurrect on reload"
+    );
+
+    // --- Container versioning POST (X-Versions-Location) ---
+    let archive_container = "photos-archive";
+    let archive_bucket = ContainerMapper::default().swift_to_s3_bucket(archive_container, project_id);
+    env.make_bucket(&archive_bucket, false).await;
+
+    container::enable_versioning(&swift_account, swift_container, archive_container, &credentials)
+        .await
+        .expect("enable versioning should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    let location = container::get_versions_location(&swift_account, swift_container, &credentials)
+        .await
+        .expect("versions location should load");
+    assert_eq!(
+        location.as_deref(),
+        Some(archive_container),
+        "versions location must survive a disk-truth metadata reload"
+    );
+
+    // --- Account metadata POST (TempURL keys etc.) ---
+    let mut account_meta = HashMap::new();
+    account_meta.insert("temp-url-key".to_string(), "s3cr3t".to_string());
+    account::update_account_metadata(&swift_account, &account_meta, &Some(credentials.clone()))
+        .await
+        .expect("account metadata POST should succeed");
+
+    reload_bucket_metadata_from_disk(&account_metadata_bucket_name(&swift_account)).await;
+
+    let loaded = account::get_account_metadata(&swift_account, &None)
+        .await
+        .expect("account metadata should load");
+    assert_eq!(
+        loaded.get("temp-url-key").map(String::as_str),
+        Some("s3cr3t"),
+        "account metadata POST must survive a disk-truth metadata reload"
+    );
+}
+
+/// The rewrites all share one tag set, so a closure that ignored the current
+/// state would still pass a single-feature test. Drive ACLs, versioning and
+/// container metadata over the same container and assert each survives the
+/// others — and that clearing one leaves the rest alone.
+async fn tag_writers_preserve_each_others_state(env: &TestECStoreEnv) {
+    let project_id = "swiftcrosstagproj";
+    let swift_account = format!("AUTH_{project_id}");
+    let credentials = keystone_credentials(project_id);
+    let container = "shared";
+    let archive = "shared-archive";
+    let bucket = ContainerMapper::default().swift_to_s3_bucket(container, project_id);
+    env.make_bucket(&bucket, false).await;
+    env.make_bucket(&ContainerMapper::default().swift_to_s3_bucket(archive, project_id), false)
+        .await;
+
+    let mut metadata = HashMap::new();
+    metadata.insert("color".to_string(), "blue".to_string());
+    update_container_metadata(&swift_account, container, &credentials, metadata)
+        .await
+        .expect("container metadata POST should succeed");
+    container::enable_versioning(&swift_account, container, archive, &credentials)
+        .await
+        .expect("enable versioning should succeed");
+    container::set_container_acl(&swift_account, container, Some(".r:*"), Some("AUTH_other"), &credentials)
+        .await
+        .expect("set container ACL should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    // All three writers' state coexists after a disk-truth reload.
+    let meta = persisted_container_metadata(&bucket).await;
+    assert_eq!(meta.get("color").map(String::as_str), Some("blue"));
+    assert_eq!(
+        container::get_versions_location(&swift_account, container, &credentials)
+            .await
+            .expect("versions location should load")
+            .as_deref(),
+        Some(archive)
+    );
+    let acl = container::get_container_acl(&swift_account, container, &credentials)
+        .await
+        .expect("container ACL should load");
+    assert!(!acl.read.is_empty(), "read ACL must survive the reload");
+    assert!(!acl.write.is_empty(), "write ACL must survive the reload");
+
+    // Disabling versioning drops only the versioning tag.
+    container::disable_versioning(&swift_account, container, &credentials)
+        .await
+        .expect("disable versioning should succeed");
+
+    reload_bucket_metadata_from_disk(&bucket).await;
+
+    assert_eq!(
+        container::get_versions_location(&swift_account, container, &credentials)
+            .await
+            .expect("versions location should load"),
+        None,
+        "disable_versioning must clear the versioning tag durably"
+    );
+    let meta = persisted_container_metadata(&bucket).await;
+    assert_eq!(
+        meta.get("color").map(String::as_str),
+        Some("blue"),
+        "disable_versioning must not disturb container metadata"
+    );
+    let acl = container::get_container_acl(&swift_account, container, &credentials)
+        .await
+        .expect("container ACL should load");
+    assert!(!acl.read.is_empty(), "disable_versioning must not disturb the ACL");
+}
+
+/// A container that does not exist must not get metadata persisted for it:
+/// the metadata loader turns "nothing on disk" into a fresh default, so an
+/// unguarded rewrite would create an orphan metadata file and cache a
+/// fabricated default as authoritative.
+async fn versioning_writes_reject_missing_containers() {
+    let project_id = "swiftmissingproj";
+    let swift_account = format!("AUTH_{project_id}");
+    let credentials = keystone_credentials(project_id);
+    let missing = "no-such-container";
+
+    let err = container::disable_versioning(&swift_account, missing, &credentials)
+        .await
+        .expect_err("disabling versioning on a missing container must fail");
+    assert!(
+        matches!(err, SwiftError::NotFound(_)),
+        "expected NotFound for a missing container, got {err:?}"
+    );
+
+    let bucket = ContainerMapper::default().swift_to_s3_bucket(missing, project_id);
+    assert!(
+        get_bucket_metadata(&bucket).await.is_err(),
+        "a rejected write must not have cached metadata for a nonexistent container"
+    );
+}
+
+/// Account metadata holds the account's TempURL signing key, and it is now
+/// durable — so a write for someone else's account would be a persistent,
+/// cluster-wide takeover of that account's pre-signed URLs, not a cache blip.
+/// The write path must reject both a foreign account and a missing token,
+/// while reads stay open for pre-auth TempURL signature validation.
+///
+/// Deliberately builds no store: both rejections must happen before the write
+/// path resolves storage at all, and a second `TestECStoreEnv` in this process
+/// would race the other test over the ambient store handle.
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn account_metadata_write_rejects_foreign_and_anonymous_callers() {
+    let victim_account = "AUTH_victimproject";
+    let attacker_credentials = keystone_credentials("attackerproject");
+
+    let mut poisoned = HashMap::new();
+    poisoned.insert("temp-url-key".to_string(), "attacker-key".to_string());
+
+    let err = account::update_account_metadata(victim_account, &poisoned, &Some(attacker_credentials))
+        .await
+        .expect_err("writing another account's metadata must be rejected");
+    assert!(
+        matches!(err, SwiftError::Forbidden(_)),
+        "cross-account metadata write must be Forbidden, got {err:?}"
+    );
+
+    let err = account::update_account_metadata(victim_account, &poisoned, &None)
+        .await
+        .expect_err("anonymous account metadata write must be rejected");
+    assert!(
+        matches!(err, SwiftError::Unauthorized(_)),
+        "anonymous metadata write must be Unauthorized, got {err:?}"
+    );
+}

rustfs/src/app/lifecycle_transition_api_test.rs

@@ -60,7 +60,6 @@ use uuid::Uuid;
 static GLOBAL_ENV: OnceLock<(Vec<PathBuf>, Arc<ECStore>)> = OnceLock::new();
 static INIT: Once = Once::new();
 const TRANSITION_WAIT_TIMEOUT: Duration = Duration::from_secs(15);
-const RESTORE_COPY_BACK_WAIT_TIMEOUT: Duration = Duration::from_secs(60);
 const ENV_GET_CODEC_STREAMING_ENABLE: &str = "RUSTFS_GET_CODEC_STREAMING_ENABLE";
 const ENV_GET_CODEC_STREAMING_ROLLOUT: &str = "RUSTFS_GET_CODEC_STREAMING_ROLLOUT";
 const ENV_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED: &str = "RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED";
@@ -339,45 +338,6 @@ async fn wait_for_transition(ecstore: &Arc<ECStore>, bucket: &str, object: &str,
     }
 }
 
-async fn wait_for_restore_completion(
-    ecstore: &Arc<ECStore>,
-    backend: &MockWarmBackend,
-    bucket: &str,
-    object: &str,
-    timeout: Duration,
-) -> Result<ObjectInfo, String> {
-    let deadline = tokio::time::Instant::now() + timeout;
-    let mut last_state = None;
-
-    loop {
-        if tokio::time::Instant::now() >= deadline {
-            let tier_gets = backend.get_count().await;
-            let op_log = backend.op_log().await;
-            return Err(format!(
-                "restore copy-back should complete within {timeout:?}; tier_gets={tier_gets}, op_log={op_log:?}; last observed state: {}",
-                last_state.unwrap_or_else(|| "no object info observed".to_string())
-            ));
-        }
-
-        match (**ecstore).get_object_info(bucket, object, &ObjectOptions::default()).await {
-            Ok(info) => {
-                if !info.restore_ongoing && info.restore_expires.is_some() {
-                    return Ok(info);
-                }
-                last_state = Some(format!(
-                    "restore_ongoing={}, restore_expires={:?}, transitioned_status={}",
-                    info.restore_ongoing, info.restore_expires, info.transitioned_object.status
-                ));
-            }
-            Err(err) => {
-                last_state = Some(format!("get_object_info failed: {err}"));
-            }
-        }
-
-        tokio::time::sleep(Duration::from_millis(500)).await;
-    }
-}
-
 // SAFETY: this helper is used only by `#[serial]` tests and runs under the single-threaded Tokio
 // runtime (`worker_threads = 1`), so no concurrent test can mutate process environment during the
 // `env::set_var` / `env::remove_var` window.
@@ -2097,10 +2057,9 @@ async fn put_bucket_lifecycle_configuration_rejects_zero_day_expiration() {
 /// POST restore(days=1) is accepted and flips the object to
 /// `x-amz-restore: ongoing-request="true"` while the mock tier GET barrier
 /// proves the background copy-back has reached the remote read; a second POST
-/// during that window is rejected with 409 `RestoreAlreadyInProgress`; once the
-/// copy-back completes the object reports `ongoing-request="false"` with a
-/// future expiry-date; and a full GET is then served from the local restored
-/// copy (the mock tier records no further `get` calls).
+/// during that window is rejected with 409 `RestoreAlreadyInProgress`.
+/// Synchronous SetDisks transition tests cover copy-back completion, restore
+/// metadata, and local byte-identical reads.
 ///
 /// Re-enabled in the serial lane by backlog#1304: the accept path now flips
 /// the ongoing flag under a short compare-and-set guard and the copy-back
@@ -2110,7 +2069,7 @@ async fn put_bucket_lifecycle_configuration_rejects_zero_day_expiration() {
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 #[serial]
 #[ignore = "global-state ILM integration test: runs serialized in the CI ILM Integration (serial) lane, see ci.yml test-ilm-integration-serial and rustfs/backlog#1148 (ilm-8)"]
-async fn restore_object_usecase_reports_ongoing_conflict_and_completion() {
+async fn restore_object_usecase_reports_ongoing_conflict() {
     let (_disk_paths, ecstore) = setup_test_env().await;
     let usecase = DefaultObjectUsecase::from_global();
 
@@ -2177,35 +2136,6 @@ async fn restore_object_usecase_reports_ongoing_conflict_and_completion() {
     );
 
     get_barrier.release();
-
-    // Completion: ongoing flips to false and a future expiry-date appears.
-    let completed =
-        wait_for_restore_completion(&ecstore, &backend, bucket.as_str(), object, RESTORE_COPY_BACK_WAIT_TIMEOUT).await;
-    let completed = completed.unwrap_or_else(|err| panic!("{err}"));
-
-    let now_secs = std::time::SystemTime::now()
-        .duration_since(std::time::UNIX_EPOCH)
-        .expect("clock before unix epoch")
-        .as_secs() as i64;
-    let expires = completed.restore_expires.expect("completed restore carries an expiry");
-    assert!(
-        expires.unix_timestamp() > now_secs,
-        "restore expiry-date must be in the future, got {expires}"
-    );
-    assert_eq!(
-        completed.transitioned_object.status, "complete",
-        "restore must not clear the transitioned state"
-    );
-
-    // The restored copy serves GET locally: no further tier GETs.
-    let tier_gets_after_restore = backend.get_count().await;
-    let data = read_object_bytes(&ecstore, bucket.as_str(), object).await;
-    assert_eq!(data, payload, "restored GET must return the original bytes");
-    assert_eq!(
-        backend.get_count().await,
-        tier_gets_after_restore,
-        "GET of a restored object must be served locally, not from the tier"
-    );
 }
 
 /// backlog#1304: the restore-accept compare-and-set itself, under real

scripts/README.md

@@ -90,7 +90,7 @@ their issue closes.
 | `manual_transition_journal_audit.sh` | dev-tool | Journal + metrics + log audit for manual transition jobs | — |
 | `manual_transition_mixed_rollout_matrix.sh` | dev-tool | Matrix generator for mixed-version rollout phases | — |
 | `manual_transition_mixed_rollout_runbook.sh` | dev-tool | Reusable mixed-version rollout runbook generator (external run) | — |
-| `manual_transition_mixed_version_docker_harness.sh` | dev-tool | Dedicated #1508 Docker harness for old/new manual-transition rollout evidence | `test_manual_transition_runbooks.sh` |
+| `manual_transition_mixed_version_docker_harness.sh` | dev-tool | Dedicated #1508 Docker harness for old/new manual-transition rollout evidence with strict/baseline/blocked result classification | `test_manual_transition_runbooks.sh` |
 | `monitor_manual_transition_ci.sh` | dev-tool | CI workflow/status watcher for manual transition follow-up monitoring | — |
 | `manual_transition_soak_matrix.sh` | dev-tool | Matrix generator for nightly stress windows | — |
 | `manual_transition_nightly_stress_runbook.sh` | dev-tool | Nightly stress entrypoint with failure snapshot templates | — |

scripts/check_layer_dependencies.sh

@@ -13,7 +13,14 @@ fi
 classify_source_layer() {
   local file="$1"
 
-  if [[ "$file" == rustfs/src/app/* ]]; then
+  if [[ "$file" == rustfs/src/server/* ]] ||
+    [[ "$file" == rustfs/src/startup_*.rs ]] ||
+    [[ "$file" == rustfs/src/init.rs ]] ||
+    [[ "$file" == rustfs/src/main.rs ]] ||
+    [[ "$file" == rustfs/src/lib.rs ]] ||
+    [[ "$file" == rustfs/src/embedded.rs ]]; then
+    printf 'composition'
+  elif [[ "$file" == rustfs/src/app/* ]]; then
     printf 'app'
   elif [[ "$file" == rustfs/src/admin/* ]] || [[ "$file" == rustfs/src/storage/ecfs.rs ]] || [[ "$file" == rustfs/src/storage/s3_api/* ]]; then
     printf 'interface'
@@ -30,6 +37,14 @@ classify_target_layer() {
   local storage_path
 
   case "$root" in
+    init | main | lib | embedded | startup_*)
+      printf 'composition'
+      ;;
+    server)
+      # Server files are composition roots when they import lower layers, but
+      # their exported HTTP contracts belong to the interface boundary.
+      printf 'interface'
+      ;;
     app)
       printf 'app'
       ;;
@@ -53,6 +68,9 @@ classify_target_layer() {
 
 layer_rank() {
   case "$1" in
+    composition)
+      printf '4'
+      ;;
     interface)
       printf '3'
       ;;
@@ -68,6 +86,48 @@ layer_rank() {
   esac
 }
 
+is_reverse_dependency() {
+  local source_rank target_rank
+
+  source_rank="$(layer_rank "$1")"
+  target_rank="$(layer_rank "$2")"
+  (( source_rank < target_rank ))
+}
+
+assert_dependency_direction() {
+  local expected="$1"
+  local source="$2"
+  local target="$3"
+  local actual='allowed'
+
+  if is_reverse_dependency "$source" "$target"; then
+    actual='reverse'
+  fi
+  if [[ "$actual" != "$expected" ]]; then
+    printf 'Layer dependency guard self-test failed: %s -> %s (expected %s, got %s)\n' \
+      "$source" "$target" "$expected" "$actual" >&2
+    exit 1
+  fi
+}
+
+run_layer_model_self_tests() {
+  local server_source app_source storage_source server_target admin_target app_target
+
+  server_source="$(classify_source_layer rustfs/src/server/http.rs)"
+  app_source="$(classify_source_layer rustfs/src/app/bucket_usecase.rs)"
+  storage_source="$(classify_source_layer rustfs/src/storage/rpc/node_service.rs)"
+  server_target="$(classify_target_layer server::http)"
+  admin_target="$(classify_target_layer admin::router)"
+  app_target="$(classify_target_layer app::bucket_usecase)"
+
+  assert_dependency_direction 'allowed' "$server_source" "$admin_target"
+  assert_dependency_direction 'allowed' "$server_source" "$app_target"
+  assert_dependency_direction 'reverse' "$app_source" "$server_target"
+  assert_dependency_direction 'reverse' "$storage_source" "$server_target"
+  assert_dependency_direction 'reverse' "$app_source" "$admin_target"
+  assert_dependency_direction 'reverse' "$storage_source" "$admin_target"
+}
+
 normalize_import_group_item() {
   local prefix="$1"
   local item="$2"
@@ -182,6 +242,11 @@ normalize_import_path() {
 
 emit_crate_use_statements() {
   (cd "$ROOT_DIR" && rg --files -g '*.rs' rustfs/src | while IFS= read -r file; do
+    # The guard is file-scoped: dedicated test modules are excluded, while
+    # inline #[cfg(test)] imports remain subject to the source file's layer.
+    if [[ "$file" == *_test.rs ]] || [[ "$file" == */tests/* ]]; then
+      continue
+    fi
     perl -0777 -ne '
       while (/\buse\s+crate::.*?;/sg) {
         my $statement = $&;
@@ -194,6 +259,26 @@ emit_crate_use_statements() {
   done)
 }
 
+write_baseline_file() {
+  local entries="$1"
+
+  cat >"$BASELINE_FILE" <<'EOF'
+# Layer dependency baseline for the rustfs binary crate.
+#
+# The guard models production imports as:
+#   composition -> interface -> app -> infra
+#
+# Canonical dependency entry:
+#   dep|source_file|source_layer->target_layer|crate::imported_symbol
+#
+# Canonical conceptual cycle entry:
+#   cycle|left_layer<->right_layer
+EOF
+  cat "$entries" >>"$BASELINE_FILE"
+}
+
+run_layer_model_self_tests
+
 normalize_baseline_file() {
   local input="$1"
   local output="$2"
@@ -265,10 +350,7 @@ while IFS= read -r line; do
       printf '%s->%s\n' "$source_layer" "$target_layer" >>"$EDGES_RAW"
     fi
 
-    source_rank="$(layer_rank "$source_layer")"
-    target_rank="$(layer_rank "$target_layer")"
-
-    if (( source_rank < target_rank )); then
+    if is_reverse_dependency "$source_layer" "$target_layer"; then
       printf 'dep|%s|%s->%s|crate::%s\n' "$file" "$source_layer" "$target_layer" "$import_path" >>"$VIOLATIONS_RAW"
     fi
   done < <(normalize_import_path "$text")
@@ -292,7 +374,7 @@ done <"${TMP_DIR}/edges_sorted.txt" | sort -u >"${TMP_DIR}/cycles_sorted.txt"
 cat "${TMP_DIR}/violations_sorted.txt" "${TMP_DIR}/cycles_sorted.txt" | sort -u >"$CURRENT_BASELINE"
 
 if [[ "$MODE" == "update" ]]; then
-  cp "$CURRENT_BASELINE" "$BASELINE_FILE"
+  write_baseline_file "$CURRENT_BASELINE"
   echo "Updated baseline: $BASELINE_FILE"
   exit 0
 fi

scripts/layer-dependency-baseline.txt

@@ -1,35 +1,44 @@
 # Layer dependency baseline for the rustfs binary crate.
 #
-# These are intra-crate module references within rustfs/ that cross the
-# conceptual layer boundaries (app, infra/server, interface/admin).
-# Since they live inside one Cargo crate, Rust doesn't enforce separation.
-# The list is maintained for architectural awareness during code review.
+# The guard models production imports as:
+#   composition -> interface -> app -> infra
 #
-# Format for dependency entries:
-#   status|source_file|direction|imported_symbol|reason
+# Canonical dependency entry:
+#   dep|source_file|source_layer->target_layer|crate::imported_symbol
 #
-# Format for conceptual cycles:
-#   status|cycle|direction_pair|reason
-#
-# Status:
-#   accepted - reviewed and intentionally allowed
-#   todo     - should be resolved in a future refactor
-
+# Canonical conceptual cycle entry:
+#   cycle|left_layer<->right_layer
 cycle|app<->infra
 cycle|app<->interface
+cycle|composition<->infra
+cycle|composition<->interface
 cycle|infra<->interface
+dep|rustfs/src/admin/handlers/scanner.rs|interface->composition|crate::startup_background::ENV_SCANNER_ENABLED
+dep|rustfs/src/admin/handlers/scanner.rs|interface->composition|crate::startup_background::scanner_enabled_from_env
+dep|rustfs/src/app/admin_usecase.rs|app->interface|crate::server::DependencyReadiness
+dep|rustfs/src/app/admin_usecase.rs|app->interface|crate::server::collect_dependency_readiness_report
 dep|rustfs/src/app/bucket_usecase.rs|app->interface|crate::admin::handlers::site_replication::site_replication_bucket_meta_hook
 dep|rustfs/src/app/bucket_usecase.rs|app->interface|crate::admin::handlers::site_replication::site_replication_delete_bucket_hook
 dep|rustfs/src/app/bucket_usecase.rs|app->interface|crate::admin::handlers::site_replication::site_replication_make_bucket_hook
-dep|rustfs/src/init.rs|infra->interface|crate::admin
+dep|rustfs/src/app/bucket_usecase.rs|app->interface|crate::server::RemoteAddr
+dep|rustfs/src/app/object_usecase.rs|app->interface|crate::server::convert_ecstore_object_info
+dep|rustfs/src/cluster_snapshot.rs|infra->interface|crate::server::DependencyReadiness
+dep|rustfs/src/cluster_snapshot.rs|infra->interface|crate::server::DependencyReadinessReport
+dep|rustfs/src/cluster_snapshot.rs|infra->interface|crate::server::ReadinessDegradedReason
+dep|rustfs/src/cluster_snapshot.rs|infra->interface|crate::server::snapshot_dependency_readiness_report
 dep|rustfs/src/runtime_sources.rs|infra->app|crate::app::context
-dep|rustfs/src/server/http.rs|infra->interface|crate::admin
-dep|rustfs/src/server/layer.rs|infra->interface|crate::admin::console::is_console_path
+dep|rustfs/src/storage/access.rs|infra->interface|crate::server::RemoteAddr
+dep|rustfs/src/storage/ecfs_extend.rs|infra->interface|crate::server::cors
 dep|rustfs/src/storage/ecfs_extend.rs|infra->interface|crate::storage::ecfs::ListObjectUnorderedQuery
-dep|rustfs/src/storage/ecfs_test.rs|infra->interface|crate::storage::ecfs::FS
-dep|rustfs/src/storage/ecfs_test.rs|infra->interface|crate::storage::ecfs::validate_object_lock_configuration_input
-dep|rustfs/src/storage/ecfs_test.rs|infra->interface|crate::storage::s3_api::common::rustfs_initiator
-dep|rustfs/src/storage/ecfs_test.rs|infra->interface|crate::storage::s3_api::common::rustfs_owner
+dep|rustfs/src/storage/helper.rs|infra->interface|crate::server::convert_ecstore_object_info
+dep|rustfs/src/storage/helper.rs|infra->interface|crate::server::is_audit_module_enabled
+dep|rustfs/src/storage/helper.rs|infra->interface|crate::server::is_notify_module_enabled
+dep|rustfs/src/storage/helper.rs|infra->interface|crate::server::refresh_audit_module_enabled
+dep|rustfs/src/storage/helper.rs|infra->interface|crate::server::refresh_notify_module_enabled
+dep|rustfs/src/storage/rpc/http_service.rs|infra->interface|crate::server::RPC_PREFIX
 dep|rustfs/src/storage/rpc/node_service.rs|infra->interface|crate::admin::service::config::reload_dynamic_config_runtime_state
 dep|rustfs/src/storage/rpc/node_service.rs|infra->interface|crate::admin::service::config::reload_runtime_config_snapshot
 dep|rustfs/src/storage/rpc/node_service.rs|infra->interface|crate::admin::service::site_replication::reload_site_replication_runtime_state
+dep|rustfs/src/storage/rpc/node_service.rs|infra->interface|crate::server::MODULE_SWITCHES_SIGNAL_SUBSYSTEM
+dep|rustfs/src/storage/rpc/node_service/heal.rs|infra->composition|crate::startup_background::heal_enabled_from_env
+dep|rustfs/src/storage/rpc/node_service/heal.rs|infra->composition|crate::startup_background::scanner_enabled_from_env

scripts/manual_transition_mixed_version_docker_harness.sh

@@ -25,9 +25,13 @@ COLD_IMAGE="${COLD_IMAGE:-${NEW_IMAGE}}"
 BASE_PORT="${BASE_PORT:-19400}"
 OUT_DIR="${OUT_DIR:-${PROJECT_ROOT}/target/manual-transition-1508-docker/$(date +%Y%m%dT%H%M%S)}"
 KEEP_UP=false
-ROLLBACK_NEW2_TO_OLD=true
+ROLLBACK_PHASE="${ROLLBACK_PHASE:-after-terminal}"
+OLD_NODE_PHASE="${OLD_NODE_PHASE:-initial}"
 WAIT_TIMEOUT_SECS="${WAIT_TIMEOUT_SECS:-180}"
 POLL_SECONDS="${POLL_SECONDS:-180}"
+TRANSITION_WORKERS="${TRANSITION_WORKERS:-2}"
+TRANSITION_QUEUE_CAPACITY="${TRANSITION_QUEUE_CAPACITY:-64}"
+FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT="${FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT:-false}"
 
 HOT_ACCESS_KEY="${HOT_ACCESS_KEY:-mvadmin}"
 HOT_SECRET_KEY="${HOT_SECRET_KEY:-mvsecret}"
@@ -58,18 +62,31 @@ Options:
   --object-count <n>           Non-empty probe object count
   --tier <name>                Remote tier name
   --keep-up                    Leave Docker services running
-  --no-rollback                Do not replace node2 with old image after job admission
+  --rollback-phase <phase>     Rollback node2 timing: after-terminal, in-flight, none
+  --old-node-phase <phase>     Old node1 timing: initial, before-job
+  --no-rollback                Alias for --rollback-phase none
   -h, --help                   Show help
 
 Environment:
   PROJECT_NAME OLD_IMAGE NEW_IMAGE COLD_IMAGE BASE_PORT OUT_DIR KEEP_UP
   HOT_ACCESS_KEY HOT_SECRET_KEY COLD_ACCESS_KEY COLD_SECRET_KEY
   TIER_NAME TIER_BUCKET TIER_PREFIX JOB_BUCKET JOB_PREFIX OBJECT_COUNT
-  WAIT_TIMEOUT_SECS POLL_SECONDS AWS_SIGV4_SCOPE
+  ROLLBACK_PHASE WAIT_TIMEOUT_SECS POLL_SECONDS TRANSITION_WORKERS
+  TRANSITION_QUEUE_CAPACITY OLD_NODE_PHASE FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT AWS_SIGV4_SCOPE
 
 Artifacts:
   compose.yml, image inspect files, health/readiness logs, API responses,
   terminal status, old-node readback, container logs, summary.env.
+
+Result classifications:
+  strict_mixed_rollout_pass          Real old/new images, non-empty completed transition, zero failures
+  baseline_tiered_storage_pass      Same old/new image completed transition; useful baseline, not #1508 closure
+  blocked_manual_api_not_implemented Manual transition API returned 501 before job admission
+  blocked_manual_api_unavailable    Manual transition API did not return a usable job_id
+  blocked_cluster_readiness_failed  Docker cluster did not reach health/readiness before admission
+  blocked_empty_scan_or_lifecycle   Job completed without lifecycle-matching transition work
+  blocked_manual_job_preempted_by_lifecycle_queue Lifecycle/immediate transition queued work before the job
+  strict_mixed_rollout_fail         Mixed rollout ran but did not satisfy the strict #1508 gate
 USAGE
 }
 
@@ -111,6 +128,28 @@ parse_positive_int() {
   fi
 }
 
+validate_rollback_phase() {
+  case "$ROLLBACK_PHASE" in
+    after-terminal|in-flight|none)
+      ;;
+    *)
+      log_error "--rollback-phase must be one of: after-terminal, in-flight, none"
+      exit 1
+      ;;
+  esac
+}
+
+validate_old_node_phase() {
+  case "$OLD_NODE_PHASE" in
+    initial|before-job)
+      ;;
+    *)
+      log_error "--old-node-phase must be one of: initial, before-job"
+      exit 1
+      ;;
+  esac
+}
+
 parse_args() {
   while [[ $# -gt 0 ]]; do
     case "$1" in
@@ -150,8 +189,16 @@ parse_args() {
         KEEP_UP=true
         shift
         ;;
+      --rollback-phase)
+        ROLLBACK_PHASE="$(arg_value "$1" "${2:-}")"
+        shift 2
+        ;;
+      --old-node-phase)
+        OLD_NODE_PHASE="$(arg_value "$1" "${2:-}")"
+        shift 2
+        ;;
       --no-rollback)
-        ROLLBACK_NEW2_TO_OLD=false
+        ROLLBACK_PHASE=none
         shift
         ;;
       -h|--help)
@@ -198,12 +245,18 @@ cleanup() {
 }
 
 write_compose_file() {
+  local node1_image
   COMPOSE_FILE="${OUT_DIR}/compose.yml"
+  node1_image="$OLD_IMAGE"
+  if [[ "$OLD_NODE_PHASE" == "before-job" ]]; then
+    node1_image="$NEW_IMAGE"
+  fi
   cat >"$COMPOSE_FILE" <<EOF
 services:
   cold:
     image: ${COLD_IMAGE}
     hostname: cold
+    user: "0:0"
     environment:
       - RUSTFS_ADDRESS=:9000
       - RUSTFS_ACCESS_KEY=${COLD_ACCESS_KEY}
@@ -222,14 +275,20 @@ services:
       - rustfs-1508-net
 
   node1:
-    image: ${OLD_IMAGE}
+    image: ${node1_image}
     hostname: node1
+    user: "0:0"
     environment: &hot-env
       - RUSTFS_ADDRESS=:9000
       - RUSTFS_ACCESS_KEY=${HOT_ACCESS_KEY}
       - RUSTFS_SECRET_KEY=${HOT_SECRET_KEY}
       - RUSTFS_VOLUMES=$(hot_volumes)
       - RUSTFS_SCANNER_ENABLED=false
+      - RUSTFS_SCANNER_CYCLE=3600
+      - RUSTFS_SCANNER_START_DELAY_SECS=3600
+      - RUSTFS_MAX_TRANSITION_WORKERS=${TRANSITION_WORKERS}
+      - RUSTFS_TRANSITION_QUEUE_CAPACITY=${TRANSITION_QUEUE_CAPACITY}
+      - RUSTFS_TEST_FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT=${FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT}
       - RUSTFS_UNSAFE_BYPASS_DISK_CHECK=true
       - RUSTFS_OBS_LOGGER_LEVEL=warn
     volumes:
@@ -245,6 +304,7 @@ services:
   node2:
     image: ${NEW_IMAGE}
     hostname: node2
+    user: "0:0"
     environment: *hot-env
     volumes:
       - node2_data_0:/data/rustfs0
@@ -259,6 +319,7 @@ services:
   node3:
     image: ${NEW_IMAGE}
     hostname: node3
+    user: "0:0"
     environment: *hot-env
     volumes:
       - node3_data_0:/data/rustfs0
@@ -273,6 +334,7 @@ services:
   node4:
     image: ${NEW_IMAGE}
     hostname: node4
+    user: "0:0"
     environment: *hot-env
     volumes:
       - node4_data_0:/data/rustfs0
@@ -433,16 +495,19 @@ seed_objects() {
 }
 
 start_transition_job() {
-  local query response job_id
+  local query response job_id run_http_code
   query="bucket=$(url_encode "$JOB_BUCKET")"
   query="${query}&prefix=$(url_encode "${JOB_PREFIX}/")"
   query="${query}&tier=$(url_encode "$TIER_NAME")"
   query="${query}&dryRun=false&maxObjects=${OBJECT_COUNT}&mode=async"
-  response="$(curl_hot POST "$(hot_endpoint 2)/rustfs/admin/v3/ilm/transition/run?${query}")"
-  printf '%s\n' "$response" >"${OUT_DIR}/run-response.json"
+  curl_hot POST "$(hot_endpoint 2)/rustfs/admin/v3/ilm/transition/run?${query}" \
+    -o "${OUT_DIR}/run-response.json" \
+    -w "%{http_code}\n" >"${OUT_DIR}/run-response.http_code" || true
+  response="$(cat "${OUT_DIR}/run-response.json" 2>/dev/null || true)"
+  run_http_code="$(cat "${OUT_DIR}/run-response.http_code" 2>/dev/null || true)"
   job_id="$(printf '%s' "$response" | jq -r '.job_id // empty')"
   if [[ -z "$job_id" ]]; then
-    log_error "manual transition response omitted job_id"
+    log_warn "manual transition response omitted job_id, http_code=${run_http_code:-unknown}"
     return 1
   fi
   printf '%s\n' "$job_id" >"${OUT_DIR}/job-id.txt"
@@ -451,19 +516,25 @@ start_transition_job() {
 replace_node2_with_old_image() {
   local network
   network="$(network_name)"
-  log_info "Replacing node2 with old image ${OLD_IMAGE} for in-flight rollback readback"
+  log_info "Replacing node2 with old image ${OLD_IMAGE} for ${ROLLBACK_PHASE} rollback readback"
   compose stop node2 >/dev/null
   compose rm -f node2 >/dev/null
   docker run -d \
     --name "${PROJECT_NAME}-node2-rollback-old" \
     --network "$network" \
     --network-alias node2 \
+    --user 0:0 \
     -p "$((BASE_PORT + 2)):9000" \
     -e RUSTFS_ADDRESS=:9000 \
     -e RUSTFS_ACCESS_KEY="$HOT_ACCESS_KEY" \
     -e RUSTFS_SECRET_KEY="$HOT_SECRET_KEY" \
     -e RUSTFS_VOLUMES="$(hot_volumes)" \
     -e RUSTFS_SCANNER_ENABLED=false \
+    -e RUSTFS_SCANNER_CYCLE=3600 \
+    -e RUSTFS_SCANNER_START_DELAY_SECS=3600 \
+    -e RUSTFS_MAX_TRANSITION_WORKERS="$TRANSITION_WORKERS" \
+    -e RUSTFS_TRANSITION_QUEUE_CAPACITY="$TRANSITION_QUEUE_CAPACITY" \
+    -e RUSTFS_TEST_FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT="$FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT" \
     -e RUSTFS_UNSAFE_BYPASS_DISK_CHECK=true \
     -e RUSTFS_OBS_LOGGER_LEVEL=warn \
     -v "${PROJECT_NAME}_node2_data_0:/data/rustfs0" \
@@ -473,6 +544,37 @@ replace_node2_with_old_image() {
     "$OLD_IMAGE" >/dev/null
 }
 
+replace_node1_with_old_image() {
+  local network
+  network="$(network_name)"
+  log_info "Replacing node1 with old image ${OLD_IMAGE} before manual transition job"
+  compose stop node1 >/dev/null
+  compose rm -f node1 >/dev/null
+  docker run -d \
+    --name "${PROJECT_NAME}-node1-before-job-old" \
+    --network "$network" \
+    --network-alias node1 \
+    --user 0:0 \
+    -p "$((BASE_PORT + 1)):9000" \
+    -e RUSTFS_ADDRESS=:9000 \
+    -e RUSTFS_ACCESS_KEY="$HOT_ACCESS_KEY" \
+    -e RUSTFS_SECRET_KEY="$HOT_SECRET_KEY" \
+    -e RUSTFS_VOLUMES="$(hot_volumes)" \
+    -e RUSTFS_SCANNER_ENABLED=false \
+    -e RUSTFS_SCANNER_CYCLE=3600 \
+    -e RUSTFS_SCANNER_START_DELAY_SECS=3600 \
+    -e RUSTFS_MAX_TRANSITION_WORKERS="$TRANSITION_WORKERS" \
+    -e RUSTFS_TRANSITION_QUEUE_CAPACITY="$TRANSITION_QUEUE_CAPACITY" \
+    -e RUSTFS_TEST_FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT="$FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT" \
+    -e RUSTFS_UNSAFE_BYPASS_DISK_CHECK=true \
+    -e RUSTFS_OBS_LOGGER_LEVEL=warn \
+    -v "${PROJECT_NAME}_node1_data_0:/data/rustfs0" \
+    -v "${PROJECT_NAME}_node1_data_1:/data/rustfs1" \
+    -v "${PROJECT_NAME}_node1_data_2:/data/rustfs2" \
+    -v "${PROJECT_NAME}_node1_data_3:/data/rustfs3" \
+    "$OLD_IMAGE" >/dev/null
+}
+
 poll_terminal_status() {
   local job_id="$1"
   local status_url status_json terminal_state
@@ -512,6 +614,85 @@ head_probe() {
     -w "%{http_code}\n" >"${OUT_DIR}/head-object.http_code" || true
 }
 
+image_id() {
+  local file="$1"
+  jq -r '.[0].Id // ""' "$file" 2>/dev/null || true
+}
+
+is_compat_readback_code() {
+  case "$1" in
+    200|501)
+      return 0
+      ;;
+    *)
+      return 1
+      ;;
+  esac
+}
+
+classify_result() {
+  local terminal_state="$1"
+  local transition_completed="$2"
+  local transition_failed="$3"
+  local tier_failure="$4"
+  local old_code="$5"
+  local rollback_code="$6"
+  local run_http_code="$7"
+  local lifecycle_config_found="$8"
+  local scanned="$9"
+  local eligible="${10}"
+  local skipped_already_transitioned="${11}"
+  local skipped_already_in_flight="${12}"
+  local old_image_id new_image_id images_are_mixed readback_ok
+
+  if [[ -f "${OUT_DIR}/readiness-failed" ]]; then
+    printf 'blocked_cluster_readiness_failed\n'
+    return
+  fi
+
+  old_image_id="$(image_id "${OUT_DIR}/old-image.inspect.json")"
+  new_image_id="$(image_id "${OUT_DIR}/new-image.inspect.json")"
+  images_are_mixed=false
+  if [[ "$OLD_IMAGE" != "$NEW_IMAGE" && -n "$old_image_id" && -n "$new_image_id" && "$old_image_id" != "$new_image_id" ]]; then
+    images_are_mixed=true
+  fi
+
+  readback_ok=false
+  if is_compat_readback_code "$old_code"; then
+    if [[ "$ROLLBACK_PHASE" == "none" ]] || is_compat_readback_code "$rollback_code"; then
+      readback_ok=true
+    fi
+  fi
+
+  if [[ "$run_http_code" == "501" ]]; then
+    printf 'blocked_manual_api_not_implemented\n'
+    return
+  fi
+  if [[ -z "$(cat "${OUT_DIR}/job-id.txt" 2>/dev/null || true)" ]]; then
+    printf 'blocked_manual_api_unavailable\n'
+    return
+  fi
+  if [[ "$terminal_state" == "completed" && ( "$lifecycle_config_found" != "true" || "$scanned" == "0" || "$eligible" == "0" || "$transition_completed" == "0" ) ]]; then
+    printf 'blocked_empty_scan_or_lifecycle\n'
+    return
+  fi
+  if [[ "$transition_completed" == "0" && "$eligible" != "0" && "$transition_failed" == "0" && "$tier_failure" == "0" ]]; then
+    if [[ "$skipped_already_transitioned" != "0" || "$skipped_already_in_flight" != "0" ]]; then
+      printf 'blocked_manual_job_preempted_by_lifecycle_queue\n'
+      return
+    fi
+  fi
+  if [[ "$terminal_state" == "completed" && "$transition_completed" != "0" && "$transition_failed" == "0" && "$tier_failure" == "0" ]]; then
+    if [[ "$images_are_mixed" == "true" && "$readback_ok" == "true" ]]; then
+      printf 'strict_mixed_rollout_pass\n'
+      return
+    fi
+    printf 'baseline_tiered_storage_pass\n'
+    return
+  fi
+  printf 'strict_mixed_rollout_fail\n'
+}
+
 collect_logs() {
   local service
   if [[ -z "$COMPOSE_FILE" || ! -f "$COMPOSE_FILE" ]]; then
@@ -520,37 +701,63 @@ collect_logs() {
   for service in cold node1 node2 node3 node4; do
     compose logs --no-color "$service" >"${OUT_DIR}/${service}.log" 2>/dev/null || true
   done
+  docker logs "${PROJECT_NAME}-node1-before-job-old" >"${OUT_DIR}/node1-before-job-old.log" 2>/dev/null || true
   docker logs "${PROJECT_NAME}-node2-rollback-old" >"${OUT_DIR}/node2-rollback-old.log" 2>/dev/null || true
 }
 
 summarize() {
-  local terminal_state transition_completed tier_failure transition_failed old_code rollback_code
+  local terminal_state transition_completed tier_failure transition_failed old_code rollback_code run_http_code lifecycle_config_found scanned eligible skipped_already_transitioned skipped_already_in_flight queue_queued queue_active result_classification
   terminal_state="$(cat "${OUT_DIR}/terminal-state.txt" 2>/dev/null || true)"
   transition_completed="$(jq -r '.report.transition_completed // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
   tier_failure="$(jq -r '.report.tier_failure // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
   transition_failed="$(jq -r '.report.transition_failed // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
   old_code="$(cat "${OUT_DIR}/old-node-status.http_code" 2>/dev/null || true)"
   rollback_code="$(cat "${OUT_DIR}/rollback-node2-status.http_code" 2>/dev/null || true)"
+  run_http_code="$(cat "${OUT_DIR}/run-response.http_code" 2>/dev/null || true)"
+  lifecycle_config_found="$(jq -r '.report.lifecycle_config_found // false' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf 'false')"
+  scanned="$(jq -r '.report.scanned // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  eligible="$(jq -r '.report.eligible // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  skipped_already_transitioned="$(jq -r '.report.skipped_already_transitioned // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  skipped_already_in_flight="$(jq -r '.report.skipped_already_in_flight // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  queue_queued="$(jq -r '.queue_snapshot.queued // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  queue_active="$(jq -r '.queue_snapshot.active // 0' "${OUT_DIR}/status-terminal.json" 2>/dev/null || printf '0')"
+  result_classification="$(classify_result "$terminal_state" "$transition_completed" "$transition_failed" "$tier_failure" "$old_code" "$rollback_code" "$run_http_code" "$lifecycle_config_found" "$scanned" "$eligible" "$skipped_already_transitioned" "$skipped_already_in_flight")"
   cat >"${OUT_DIR}/summary.env" <<EOF
 project_name=${PROJECT_NAME}
 old_image=${OLD_IMAGE}
 new_image=${NEW_IMAGE}
 cold_image=${COLD_IMAGE}
+old_image_id=$(image_id "${OUT_DIR}/old-image.inspect.json")
+new_image_id=$(image_id "${OUT_DIR}/new-image.inspect.json")
 base_port=${BASE_PORT}
 tier=${TIER_NAME}
 job_bucket=${JOB_BUCKET}
 job_prefix=${JOB_PREFIX}
 object_count=${OBJECT_COUNT}
+transition_workers=${TRANSITION_WORKERS}
+transition_queue_capacity=${TRANSITION_QUEUE_CAPACITY}
+force_immediate_transition_enqueue_timeout=${FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT}
+run_http_code=${run_http_code}
 terminal_state=${terminal_state}
+lifecycle_config_found=${lifecycle_config_found}
+scanned=${scanned}
+eligible=${eligible}
+skipped_already_transitioned=${skipped_already_transitioned}
+skipped_already_in_flight=${skipped_already_in_flight}
 transition_completed=${transition_completed}
 transition_failed=${transition_failed}
 tier_failure=${tier_failure}
+queue_queued=${queue_queued}
+queue_active=${queue_active}
 old_node_status_http_code=${old_code}
 rollback_node2_status_http_code=${rollback_code}
-rollback_new2_to_old=${ROLLBACK_NEW2_TO_OLD}
+rollback_phase=${ROLLBACK_PHASE}
+old_node_phase=${OLD_NODE_PHASE}
+rollback_new2_to_old=$([[ "$ROLLBACK_PHASE" == "none" ]] && printf 'false' || printf 'true')
+result_classification=${result_classification}
 EOF
   cat "${OUT_DIR}/summary.env"
-  if [[ "$terminal_state" != "completed" || "$transition_completed" == "0" || "$transition_failed" != "0" || "$tier_failure" != "0" ]]; then
+  if [[ "$result_classification" != "strict_mixed_rollout_pass" ]]; then
     log_error "strict #1508 mixed-version transition gate failed; see ${OUT_DIR}"
     return 1
   fi
@@ -560,6 +767,8 @@ main() {
   parse_args "$@"
   parse_positive_int "--base-port" "$BASE_PORT"
   parse_positive_int "--object-count" "$OBJECT_COUNT"
+  validate_rollback_phase
+  validate_old_node_phase
   require_cmd docker
   require_cmd curl
   require_cmd jq
@@ -574,20 +783,37 @@ main() {
   docker image inspect "$COLD_IMAGE" >"${OUT_DIR}/cold-image.inspect.json"
 
   compose up -d
-  wait_cluster_ready
+  if ! wait_cluster_ready; then
+    printf 'cluster readiness failed before manual transition admission\n' >"${OUT_DIR}/readiness-failed"
+    collect_logs
+    summarize
+    return 1
+  fi
 
   curl_cold PUT "$(cold_endpoint)/${TIER_BUCKET}" -o "${OUT_DIR}/create-cold-bucket.response" -w "%{http_code}\n" >"${OUT_DIR}/create-cold-bucket.http_code"
   create_bucket "$(hot_endpoint 2)" "$JOB_BUCKET" "${OUT_DIR}/create-hot-bucket"
   add_tier
-  put_lifecycle
   seed_objects
-  start_transition_job
-  if [[ "$ROLLBACK_NEW2_TO_OLD" == "true" ]]; then
-    replace_node2_with_old_image
+  put_lifecycle
+  if [[ "$OLD_NODE_PHASE" == "before-job" ]]; then
+    replace_node1_with_old_image
+    wait_http_ok "$(hot_endpoint 1)/health" "node1-old-live" || true
+    wait_http_ok "$(hot_endpoint 1)/health/ready" "node1-old-ready" || true
+  fi
+  if start_transition_job; then
+    if [[ "$ROLLBACK_PHASE" == "in-flight" ]]; then
+      replace_node2_with_old_image
+    fi
+    poll_terminal_status "$(cat "${OUT_DIR}/job-id.txt")" || true
+    if [[ "$ROLLBACK_PHASE" == "after-terminal" ]]; then
+      replace_node2_with_old_image
+      wait_http_ok "$(hot_endpoint 2)/health" "rollback-node2-live" || true
+    fi
+    capture_old_node_readback "$(cat "${OUT_DIR}/job-id.txt")"
+    head_probe
+  else
+    log_warn "Skipping terminal polling because no manual transition job was admitted"
   fi
-  poll_terminal_status "$(cat "${OUT_DIR}/job-id.txt")"
-  capture_old_node_readback "$(cat "${OUT_DIR}/job-id.txt")"
-  head_probe
   collect_logs
   summarize
 }

scripts/test_manual_transition_runbooks.sh

@@ -208,7 +208,16 @@ bash "$MIXED_DOCKER_HARNESS" --help >/tmp/manual_transition_mixed_version_docker
 rg -q "mixed_version_docker_harness" /tmp/manual_transition_mixed_version_docker_harness.help
 rg -q -- "--old-image" /tmp/manual_transition_mixed_version_docker_harness.help
 rg -q -- "--new-image" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q -- "--rollback-phase" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q -- "--old-node-phase" /tmp/manual_transition_mixed_version_docker_harness.help
 rg -q -- "--no-rollback" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "strict_mixed_rollout_pass" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "baseline_tiered_storage_pass" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "blocked_manual_api_not_implemented" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "blocked_cluster_readiness_failed" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "blocked_manual_job_preempted_by_lifecycle_queue" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "OLD_NODE_PHASE" /tmp/manual_transition_mixed_version_docker_harness.help
+rg -q "FORCE_IMMEDIATE_TRANSITION_ENQUEUE_TIMEOUT" /tmp/manual_transition_mixed_version_docker_harness.help
 if bash "$FAILURE_SAMPLES" --endpoint http://127.0.0.1:9000 --sample >/tmp/manual_transition_failure_samples.err 2>&1; then
   echo "failure samples script should fail when --sample has no value" >&2
   exit 1