Add bf-tree direct-streaming capacity regression guardrail

The bf-tree direct (non-Managed) streaming path uses absolute runbook tag
IDs as provider slot IDs, so the provider must be sized to the dataset ID
space rather than the runbook's max concurrent active-point count. The
existing test_data runbook does not exercise this because its max active
count equals its max tag.

Add a minimal, search-free runbook whose max tag (191) exceeds its max
active-point count (128) via a delete-then-insert sequence, plus a bftree
streaming config and an integration test. With correct sizing the test
passes; a provider mis-sized to max_pts fails with "Vector id is out of
boundary in the dataset".

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: Jordan Maples

diskann-benchmark/example/graph-index-bftree-stream-capacity.json

@@ -0,0 +1,67 @@
+{
+  "search_directories": [
+    "test_data/disk_index_search"
+  ],
+  "jobs": [
+    {
+      "type": "graph-index-stream-bftree",
+      "content": {
+        "build": {
+          "data_type": "float32",
+          "data": "disk_index_siftsmall_learn_256pts_data.fbin",
+          "distance": "squared_l2",
+          "max_degree": 32,
+          "l_build": 50,
+          "start_point_strategy": "medoid",
+          "alpha": 1.2,
+          "backedge_ratio": 1.0,
+          "num_threads": 1
+        },
+        "search": {
+          "queries": "disk_index_sample_query_10pts.fbin",
+          "reps": 1,
+          "num_threads": 1,
+          "search_l": 40,
+          "search_n": 10,
+          "recall_k": 10
+        },
+        "runbook_params": {
+          "runbook_path": "stream_capacity_runbook.yaml",
+          "dataset_name": "sift-small-256",
+          "gt_directory": "example_runbook_gt",
+          "ip_delete_method": {
+            "method": "one_hop"
+          },
+          "ip_delete_num_to_replace": 3
+        },
+        "quantization": {
+          "kind": "none"
+        },
+        "vector_store_config": {
+          "cb_size_byte": 67108864,
+          "leaf_page_size": 4096,
+          "cb_max_record_size": null,
+          "cb_min_record_size": null,
+          "read_promotion_rate": null,
+          "scan_promotion_rate": null,
+          "cb_copy_on_access_ratio": null,
+          "read_record_cache": null,
+          "cache_only": null,
+          "use_snapshot": false
+        },
+        "neighbor_store_config": {
+          "cb_size_byte": 67108864,
+          "leaf_page_size": 4096,
+          "cb_max_record_size": null,
+          "cb_min_record_size": null,
+          "read_promotion_rate": null,
+          "scan_promotion_rate": null,
+          "cb_copy_on_access_ratio": null,
+          "read_record_cache": null,
+          "cache_only": null,
+          "use_snapshot": false
+        }
+      }
+    }
+  ]
+}

diskann-benchmark/src/main.rs

@@ -369,6 +369,20 @@ mod tests {
         run_integration_test(raw);
     }
 
+    // Guardrail: the bf-tree direct (non-Managed) streaming path uses absolute
+    // runbook tag IDs as provider slot IDs, so the provider must be sized to the
+    // dataset ID space rather than the runbook's max concurrent active-point
+    // count. This runbook drives the max tag above the max active count, which
+    // fails with "Vector id is out of boundary in the dataset" if the provider
+    // is mis-sized to `max_pts`.
+    #[test]
+    #[cfg(feature = "bftree")]
+    fn graph_index_bftree_stream_capacity_integration() {
+        let raw =
+            value_from_file(&example_directory().join("graph-index-bftree-stream-capacity.json"));
+        run_integration_test(raw);
+    }
+
     #[test]
     #[cfg(feature = "bftree")]
     fn graph_index_bftree_save_load_roundtrip() {

test_data/disk_index_search/stream_capacity_runbook.yaml

@@ -0,0 +1,24 @@
+# Regression guardrail for bf-tree direct (non-Managed) streaming.
+#
+# The direct path uses absolute runbook tag IDs as provider slot IDs, so the
+# provider must be sized to the dataset ID space, not the runbook's maximum
+# concurrent active-point count. This runbook deliberately drives the maximum
+# tag (191) above the maximum number of concurrently-active points (128) via a
+# delete-then-insert sequence, so a provider sized to `max_pts` would fail with
+# "Vector id is out of boundary in the dataset" on stage 3.
+#
+# There are no search stages, so no groundtruth files are required.
+sift-small-256:
+  max_pts: 128
+  1:
+    operation: "insert"
+    start: 0
+    end: 128
+  2:
+    operation: "delete"
+    start: 0
+    end: 64
+  3:
+    operation: "insert"
+    start: 128
+    end: 192