diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index bbe5893..1ea6c37 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -173,15 +173,18 @@ jobs:
           OMNIGRAPH_UPDATE_OPENAPI: ${{ (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository) && '1' || '' }}
         run: cargo test --workspace --locked
 
-      - name: Run failpoints feature test
+      - name: Run failpoints feature tests
         if: needs.classify_changes.outputs.run_full_ci == 'true'
         # Run after the workspace test so the build cache is warm —
         # enabling --features failpoints is just an incremental rebuild
-        # of omnigraph-engine + the small `fail` crate, not the full
+        # of the target crate + the small `fail` crate, not the full
         # dep tree (lance, datafusion). A separate job with its own
         # cache key would be a fresh ~20min build on first run; this
-        # is ~30s on a warm cache.
-        run: cargo test --locked -p omnigraph-engine --features failpoints --test failpoints
+        # is ~30s on a warm cache. The cluster feature does not enable
+        # omnigraph/failpoints, so each line rebuilds only its crate.
+        run: |
+          cargo test --locked -p omnigraph-engine --features failpoints --test failpoints
+          cargo test --locked -p omnigraph-cluster --features failpoints --test failpoints
 
       - name: Commit regenerated openapi.json to PR branch
         if: |
diff --git a/crates/omnigraph-cluster/tests/failpoints.rs b/crates/omnigraph-cluster/tests/failpoints.rs
index 3ede30c..05d2913 100644
--- a/crates/omnigraph-cluster/tests/failpoints.rs
+++ b/crates/omnigraph-cluster/tests/failpoints.rs
@@ -117,3 +117,102 @@ fn failpoint_wiring_returns_injected_diagnostic() {
     drop(_failpoint);
     scenario.teardown();
 }
+
+/// Crash between the payload phase and the state write: blobs are on disk,
+/// state.json is byte-identical, nothing is acknowledged — and a plain re-run
+/// repairs by trusting the existing content-addressed blobs.
+#[test]
+fn apply_crash_after_payload_phase_leaves_state_unmoved_then_recovers() {
+    let scenario = FailScenario::setup();
+    let dir = fixture();
+    let digests = seed_applyable_state(dir.path());
+    let state_before = fs::read(state_path(dir.path())).unwrap();
+
+    {
+        let _failpoint = ScopedFailPoint::new("cluster_apply.after_payload_phase", "return");
+        let out = apply_config_dir(dir.path());
+        assert!(!out.ok);
+        assert!(!out.state_written);
+        assert!(!out.converged);
+        assert_eq!(out.applied_count, 0);
+        // Persisted pre-apply snapshot: no phantom Applied statuses.
+        assert!(
+            !out.resource_statuses
+                .contains_key("query.knowledge.find_person"),
+            "{:?}",
+            out.resource_statuses
+        );
+        // State has not moved; payloads are inert on disk; the lock released.
+        assert_eq!(fs::read(state_path(dir.path())).unwrap(), state_before);
+        assert!(query_blob(dir.path(), &digests).exists());
+        assert!(!dir.path().join("__cluster/lock.json").exists());
+    }
+
+    // The repair is a plain re-run: existing blobs are trusted by digest.
+    let recovered = apply_config_dir(dir.path());
+    assert!(recovered.ok, "{:?}", recovered.diagnostics);
+    assert!(recovered.converged);
+    assert!(recovered.state_written);
+    assert_eq!(
+        recovered.resource_statuses["query.knowledge.find_person"].status,
+        omnigraph_cluster::ResourceLifecycleStatus::Applied
+    );
+    scenario.teardown();
+}
+
+/// A concurrent writer mutating state.json between apply's read and its write
+/// (possible under `state.lock: false`) must surface `state_cas_mismatch`,
+/// acknowledge nothing, and leave the concurrent writer's state on disk.
+#[test]
+fn apply_cas_race_surfaces_state_cas_mismatch() {
+    let scenario = FailScenario::setup();
+    let dir = fixture();
+    let digests = seed_applyable_state(dir.path());
+
+    // Simulate the concurrent writer at the exact race window: rewrite
+    // state.json (valid JSON, graph/schema digests preserved, revision 99)
+    // after apply read it but before apply writes.
+    let race_path = state_path(dir.path());
+    fail::cfg_callback("cluster_apply.before_state_write", move || {
+        let mut state: serde_json::Value =
+            serde_json::from_str(&fs::read_to_string(&race_path).unwrap()).unwrap();
+        state["state_revision"] = serde_json::json!(99);
+        fs::write(&race_path, serde_json::to_string_pretty(&state).unwrap()).unwrap();
+    })
+    .expect("configure callback failpoint");
+
+    let out = apply_config_dir(dir.path());
+    fail::remove("cluster_apply.before_state_write");
+
+    assert!(!out.ok);
+    assert!(!out.state_written);
+    assert!(
+        out.diagnostics
+            .iter()
+            .any(|diagnostic| diagnostic.code == "state_cas_mismatch"),
+        "{:?}",
+        out.diagnostics
+    );
+    // Persisted snapshot, not the unwritten in-memory mutations.
+    assert!(
+        !out.resource_statuses
+            .contains_key("query.knowledge.find_person")
+    );
+    // The concurrent writer's state is what's on disk; apply's mutation never landed.
+    let state: serde_json::Value =
+        serde_json::from_str(&fs::read_to_string(state_path(dir.path())).unwrap()).unwrap();
+    assert_eq!(state["state_revision"], 99);
+    assert!(
+        state["applied_revision"]["resources"]
+            .get("query.knowledge.find_person")
+            .is_none()
+    );
+    // Blobs written before the race are inert.
+    assert!(query_blob(dir.path(), &digests).exists());
+
+    // Recovery is a plain re-run against the rewritten state.
+    let recovered = apply_config_dir(dir.path());
+    assert!(recovered.ok, "{:?}", recovered.diagnostics);
+    assert!(recovered.converged);
+    scenario.teardown();
+}