diff --git a/crates/lance-context-master/src/routes.rs b/crates/lance-context-master/src/routes.rs
index c9d7cd8..15ce912 100644
--- a/crates/lance-context-master/src/routes.rs
+++ b/crates/lance-context-master/src/routes.rs
@@ -568,6 +568,34 @@ pub async fn enqueue_task(
     Ok((StatusCode::ACCEPTED, Json(task)))
 }
 
+/// Cursor for the read-only merge failure ledger.
+#[derive(Debug, Deserialize)]
+pub struct MergeFailureParams {
+    pub after: Option<String>,
+}
+
+/// Read the durable retry ledger without probing workers or opening WAL data.
+pub async fn list_merge_failures(
+    State(state): State<Arc<MasterState>>,
+    Query(params): Query<MergeFailureParams>,
+) -> Result<Json<serde_json::Value>, MasterError> {
+    let (failures, next) = state
+        .task_store
+        .merge_coordinator()
+        .failure_page(params.after.as_deref(), 256)
+        .await
+        .map_err(|error| {
+            if error == "invalid merge failure cursor" {
+                MasterError::InvalidRequest(error)
+            } else {
+                MasterError::Internal(error)
+            }
+        })?;
+    Ok(Json(
+        serde_json::json!({"failures": failures, "next": next}),
+    ))
+}
+
 /// `GET /api/v1/scheduler/cooldowns` — targets the auto-sweeps are skipping
 /// because they failed repeatedly, with the failure count, when the cooldown
 /// lapses, and the last error. A store in this list is broken in a way that
@@ -652,6 +680,7 @@ pub fn api_router() -> Router<Arc<MasterState>> {
         .route("/experiments/{name}/rescan", post(rescan_experiment))
         .route("/tasks", post(enqueue_task).get(list_tasks))
         .route("/scheduler/cooldowns", get(list_cooldowns))
+        .route("/scheduler/merge-failures", get(list_merge_failures))
         .route("/scheduler/repairs", get(list_repairs))
         .route("/tasks/{id}", get(get_task))
         .route("/rescan", post(rescan))
@@ -805,6 +834,78 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    #[ignore = "requires ETCD_TEST_ENDPOINTS"]
+    async fn merge_failure_api_pages_metadata_without_resetting_retry_budgets() {
+        use lance_context_api::TaskKind;
+        let dir = TempDir::new().unwrap();
+        let state = MasterState::new(test_config(&dir)).await.unwrap();
+        state
+            .task_store
+            .enqueue(TaskKind::MergeWal, "hot", Vec::new())
+            .await
+            .unwrap();
+        let claim = state
+            .task_store
+            .claim_next_of_kinds(crate::task_store::TaskKinds::MERGE_WAL)
+            .await
+            .unwrap()
+            .unwrap();
+        let coordinator = state.task_store.merge_coordinator();
+        let proof = state.task_store.merge_claim(&claim);
+        for i in 0..257 {
+            coordinator
+                .record_failure(&proof, "hot", &format!("worker-{i:03}"), "storage timeout")
+                .await
+                .unwrap();
+        }
+        let before = coordinator
+            .failure("hot", "worker-000")
+            .await
+            .unwrap()
+            .unwrap();
+        let Json(first) = list_merge_failures(
+            State(state.clone()),
+            Query(MergeFailureParams { after: None }),
+        )
+        .await
+        .unwrap();
+        assert_eq!(first["failures"].as_array().unwrap().len(), 256);
+        let cursor = first["next"].as_str().unwrap().to_string();
+        let Json(second) = list_merge_failures(
+            State(state.clone()),
+            Query(MergeFailureParams {
+                after: Some(cursor),
+            }),
+        )
+        .await
+        .unwrap();
+        assert_eq!(second["failures"].as_array().unwrap().len(), 1);
+        assert!(second["next"].is_null());
+        let after = coordinator
+            .failure("hot", "worker-000")
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(after.consecutive_attempts, before.consecutive_attempts);
+        assert_eq!(after.next_retry_ms, before.next_retry_ms);
+        assert!(matches!(
+            list_merge_failures(
+                State(state.clone()),
+                Query(MergeFailureParams {
+                    after: Some("/unrelated-prefix/key".into()),
+                })
+            )
+            .await,
+            Err(MasterError::InvalidRequest(_))
+        ));
+        state
+            .task_store
+            .finish(claim, Err("test complete".into()))
+            .await
+            .unwrap();
+    }
+
     /// Create N experiments via core, register them, scan, and assert the stats
     /// table + list endpoint reflect them.
     #[tokio::test]
diff --git a/docs/merge-failures.md b/docs/merge-failures.md
new file mode 100644
index 0000000..6b867b7
--- /dev/null
+++ b/docs/merge-failures.md
@@ -0,0 +1,17 @@
+# Inspecting merge failures
+
+`GET /api/v1/scheduler/merge-failures?after=<cursor>` reads the durable retry
+ledger. It returns `failures` (at most 256 records) and `next` (an opaque cursor,
+or null at the end). An invalid cursor returns HTTP 400. The endpoint only reads
+etcd metadata; it never opens WAL payloads, probes workers, retries a merge, or
+resets a failure budget. A ledger changing during pagination is not a snapshot.
+
+Each record includes `target`, `endpoint`, `class`, `consecutive_attempts`,
+`last_error`, `last_failure_ms`, `next_retry_ms`, and `needs_attention`.
+Timestamps are Unix milliseconds. Errors are capped at 4096 characters in the
+runtime ledger. `needs_attention` means automatic probes are insufficient to
+resolve the incident; it does not authorize data deletion or bypass ownership.
+
+Retry, storage-fencing, and rollout behavior are defined in
+[merge-recovery.md](merge-recovery.md). This API adds observation only and does
+not change those policies or provide an external notification integration.