fix(cluster): forwarded item/embedding writes never replicated; heal backfill 401'd under auth

Two correctness bugs in multi-process cluster mode (m8p10), both found live on a
real 3-pod k3s cluster while validating the deployment, both invisible to the
existing in-process / unauthenticated test suites:

1. Forwarded item/embedding writes dropped on the floor. create_item /
   write_embedding gated the leader's peer broadcast on `if internal { return }`.
   A write to a NON-leader gateway is forwarded to the leader with the internal
   marker set (loop-prevention), so it hit that branch and terminated WITHOUT
   broadcasting — the item landed only on the leader. Signals were unaffected
   (the WAL relay ships regardless of the marker), which masked it. Fix: gate on
   leadership, not the marker — a follower applying a marked broadcast/heal
   terminates; the leader (external OR forwarded) always fans out. Forwarded
   writes now return the {replicated_to,failed} report instead of a bodyless null.

2. Heal item/embedding backfill 401'd whenever TIDAL_API_KEY is set.
   post_marked_blocking sent the internal marker but no Authorization header. The
   marker is a trust signal, not an auth bypass (the bearer middleware runs
   first), so every backfill POST was rejected 401 — a region that missed an item
   while down stayed permanently inconsistent at lag 0. Unauthenticated tests
   never caught it. Fix: thread TIDAL_API_KEY into RegionClusterState and attach
   it (same key on every region) to the backfill POSTs.

Verified live: forwarded writes via follower gateways converge to all 3 regions;
a region scaled to 0 during an item write backfills on heal (item_failures=0).
Follow-up: add multiproc regression tests with auth for both paths.

tidal-server/src/cluster/node.rs

@@ -123,6 +123,12 @@ pub struct RegionClusterState {
     /// if a second transport (gRPC reads, native protocol) lands, extract a
     /// `ShardTransport` trait in `scatter_gather` rather than branching here.
     blocking_client: reqwest::blocking::Client,
+    /// Bearer key for this node's OWN inter-sibling POSTs (the heal item/embedding
+    /// backfill in [`Self::post_marked_blocking`]). The internal marker is a trust
+    /// signal, NOT an auth bypass — the bearer middleware still runs first — so a
+    /// backfill POST 401s unless it carries `Authorization: Bearer` when
+    /// `TIDAL_API_KEY` is set. Same key on every region (required in multi-process).
+    api_key: Option<Arc<str>>,
     /// Signal type name → `u8` WAL id (the same u8-guarded map `SimulatedCluster`
     /// builds). The relay stays schema-agnostic; the handler resolves the id.
     signal_type_ids: HashMap<String, u8>,
@@ -170,6 +176,7 @@ impl RegionClusterState {
         profiles: Vec<tidaldb::ranking::profile::RankingProfile>,
         data_dir: Option<std::path::PathBuf>,
         hlc_offset_ms: i64,
+        api_key: Option<Arc<str>>,
     ) -> Result<Self> {
         super::topology::validate_multiproc(topology, region_name)?;
 
@@ -284,6 +291,7 @@ impl RegionClusterState {
             client,
             broadcast_peer_timeout: topology.broadcast_peer_timeout(),
             blocking_client,
+            api_key,
             signal_type_ids,
             write_pool,
             shutting_down: AtomicBool::new(false),
@@ -561,7 +569,7 @@ impl RegionClusterState {
     /// the next heal). Carries the internal marker so the peer applies locally and
     /// does not re-broadcast.
     fn post_marked_blocking(&self, url: &str, body: &serde_json::Value) -> bool {
-        match self
+        let mut req = self
             .blocking_client
             .post(url)
             .timeout(super::forward::STATUS_PEER_TIMEOUT)
@@ -569,9 +577,15 @@ impl RegionClusterState {
                 super::forward::INTERNAL_MARKER,
                 super::forward::INTERNAL_MARKER_VALUE,
             )
-            .json(body)
+            .json(body);
-            .send()
+        // The internal marker is a trust signal, not an auth bypass: when the
-        {
+        // cluster runs with TIDAL_API_KEY set, the peer's bearer middleware runs
+        // first and 401s an unauthenticated backfill. Carry this node's own key
+        // (identical on every region) so the heal item/embedding backfill lands.
+        if let Some(key) = &self.api_key {
+            req = req.bearer_auth(key);
+        }
+        match req.send() {
             Ok(resp) if resp.status().is_success() => true,
             Ok(resp) => {
                 tracing::warn!(url, status = %resp.status(), "heal backfill peer non-success");
@@ -1442,14 +1456,21 @@ pub async fn create_item(
     offload_region_read(move || RegionClusterState::apply_item_local(&db, entity, &metadata))
         .await?;
 
-    if internal {
+    if !state.is_leader() {
-        // A marked broadcast terminates here: no fan-out.
+        // A FOLLOWER applying a marked broadcast/heal terminates here: no
+        // fan-out. Gating on `internal` alone was a bug: a forwarded item write
+        // reaches the LEADER with the marker set (forward_write marks it for
+        // loop-prevention) and so terminated without ever broadcasting — items
+        // written through a non-leader gateway never replicated to peers (only
+        // the WAL-relayed signal did). Gate on leadership: a follower with the
+        // marker terminates; the leader (external OR forwarded) always fans out.
         return Ok(StatusCode::CREATED.into_response());
     }
-    // Leader external request: broadcast to peers (best-effort, per-peer
+    // Leader request (external client OR a write forwarded from a follower):
-    // report). 201 asserts the LOCAL create — the single-node /items contract —
+    // broadcast to peers (best-effort, per-peer report). 201 asserts the LOCAL
-    // while the body carries the broadcast report (cf. /embeddings' 200, which
+    // create — the single-node /items contract — while the body carries the
-    // exists only because a 204 cannot carry that body).
+    // broadcast report (cf. /embeddings' 200, which exists only because a 204
+    // cannot carry that body).
     let outcome = broadcast_to_peers(&state, "/items", &req, &headers).await;
     Ok((StatusCode::CREATED, Json(broadcast_body(&outcome))).into_response())
 }
@@ -1485,12 +1506,15 @@ pub async fn write_embedding(
     offload_region_read(move || RegionClusterState::apply_embedding_local(&db, entity, &values))
         .await?;
 
-    if internal {
+    if !state.is_leader() {
+        // Follower applying a marked broadcast/heal: terminate, no fan-out. A
+        // forwarded embedding write reaches the LEADER with the marker set and
+        // must still broadcast (see create_item for the full rationale).
         return Ok(StatusCode::NO_CONTENT.into_response());
     }
-    // Leader external request: broadcast. The per-peer report needs a body, so we
+    // Leader request (external OR forwarded from a follower): broadcast. The
-    // return 200 with the report (a 204 cannot carry one). A forwarded/internal
+    // per-peer report needs a body, so we return 200 with the report (a 204
-    // write keeps the bodyless 204 above.
+    // cannot carry one). A follower's marked write keeps the bodyless 204 above.
     let outcome = broadcast_to_peers(&state, "/embeddings", &req, &headers).await;
     Ok((StatusCode::OK, Json(broadcast_body(&outcome))).into_response())
 }

tidal-server/src/main.rs

@@ -241,6 +241,12 @@ async fn run_region_cluster(args: ClusterArgs, region: String) -> Result<()> {
     );
 
     let data_dir = args.data_dir.clone();
+    // Read the bearer key up front: the region node needs its OWN copy to
+    // authenticate inter-sibling POSTs (the heal item/embedding backfill). The
+    // internal marker is a trust signal, not an auth bypass, so an unauthenticated
+    // backfill 401s when TIDAL_API_KEY is set. Same key serves the router below.
+    let api_key = read_api_key();
+    let node_key = api_key.clone();
     // `RegionClusterState::new` builds the GrpcTransport via `GrpcTransport::new`,
     // which blocks on its own tokio runtime — must run off this reactor. Build it
     // on a dedicated thread, exactly like the single-process path.
@@ -254,13 +260,13 @@ async fn run_region_cluster(args: ClusterArgs, region: String) -> Result<()> {
                 profiles,
                 data_dir,
                 hlc_offset_ms,
+                node_key,
             )
         })
         .map_err(|e| ServerError::Cluster(format!("spawn region builder thread: {e}")))?
         .join()
         .map_err(|_| ServerError::Cluster("region builder thread panicked".into()))??;
 
-    let api_key = read_api_key();
     serve_state(state, &args.listen, api_key, build_region_router).await
 }
 

tidal-server/tests/cluster_region.rs

@@ -111,7 +111,7 @@ impl Pair {
 fn build_region(topology: TopologySpec, region: &str) -> RegionClusterState {
     let region = region.to_string();
     std::thread::spawn(move || {
-        RegionClusterState::new(&topology, &region, region_schema(), Vec::new(), None, 0)
+        RegionClusterState::new(&topology, &region, region_schema(), Vec::new(), None, 0, None)
     })
     .join()
     .unwrap()

tidal-server/tests/cluster_routes.rs

@@ -102,7 +102,7 @@ impl Cluster3 {
 fn build_region(topology: TopologySpec, region: &str) -> RegionClusterState {
     let region = region.to_string();
     std::thread::spawn(move || {
-        RegionClusterState::new(&topology, &region, region_schema(), Vec::new(), None, 0)
+        RegionClusterState::new(&topology, &region, region_schema(), Vec::new(), None, 0, None)
     })
     .join()
     .unwrap()