diff --git a/notes/bf-2ynu5.md b/notes/bf-2ynu5.md
new file mode 100644
index 0000000..1b9ce82
--- /dev/null
+++ b/notes/bf-2ynu5.md
@@ -0,0 +1,78 @@
+# bf-2ynu5 — Add per-node document-count query helper for reshard backfill (umbrella close-out)
+
+Split-child of bf-49a17 (reshard backfill progress never populated). This was
+the **foundation/umbrella** piece: a reusable mechanism to compute a reshard
+source index's real document count, with no change to the reshard state machine.
+
+The implementation landed as **four closed sibling/child beads**, each owning
+one slice. This bead's own deliverable is the umbrella close-out: confirm the
+children integrate, build, and pass, and that every acceptance criterion below
+is met in the current tree. No code was authored under bf-2ynu5 directly.
+
+## Child beads (all closed)
+
+| Bead | Commit | What it owned |
+|------|--------|---------------|
+| bf-2r01e | f10a675 | Register `index_stats` module (`pub mod index_stats;` lib.rs:37) |
+| bf-14wgq | 3dd2b75 | Add `ReshardExecutor::compute_source_document_count` (executor.rs:538) |
+| bf-1s9od | 1523cae | Unit-test the executor helper against mockito nodes (executor.rs:970+) |
+| bf-3r626 | 3b52aba | Migrate ILM off its duplicate stats code onto the shared module |
+
+## Acceptance criteria — verified against current tree (2026-07-10)
+
+1. **Helper queries each healthy source node's `GET /indexes/{uid}/stats` and
+   returns an aggregated count.**
+   `compute_source_document_count` (reshard/executor.rs:538) delegates to
+   `index_stats::aggregate_index_stats` (index_stats.rs:130), which iterates the
+   source `node_addresses`, GETs `/indexes/{uid}/stats`, and reduces
+   `numberOfDocuments` with `max` (each address hosts a full replica, so `sum`
+   would over-count by the replication factor — same reduction ILM relies on).
+   A node that fails to respond (network / non-2xx) is logged and skipped; a 404
+   (absent replica) counts as zero. "Healthy" = "responds successfully", matching
+   the original `ilm::fetch_index_stats` precedent.
+
+2. **Reuses/extracts the existing ilm.rs aggregation rather than duplicating.**
+   The HTTP+reduce logic lives once in the shared `index_stats` module
+   (index_stats.rs). ILM's old `fetch_index_stats` is gone; ilm.rs now calls
+   `crate::index_stats::aggregate_index_stats` at :701 and :1067. Reshard and
+   ILM therefore compute an index's count the *same* way.
+
+3. **Unit-tested against a mock node client returning known per-node counts.**
+   - Pure reduction policy `reduce_document_counts` (index_stats.rs:170), 9 tests.
+   - Executor helper driven against mockito `GET /indexes/{uid}/stats` servers,
+     4 tests (max across healthy; one failing node skipped; 404 tolerated;
+     all-fail → 0). See "Test run" below.
+
+4. **No change to the reshard state machine / backfill migration.**
+   `start_backfill` (executor.rs:513) still initializes `total_documents: 0`
+   (`// Will be updated`); `advance_backfill` (executor.rs:640) still accumulates
+   `total_docs_in_shard` incrementally. `compute_source_document_count` is a pure
+   additive helper that is *not* called from the state machine in this umbrella.
+   Wiring the up-front denominator into `start_backfill` is intentionally a
+   sibling bead's job.
+
+## Test run (this close-out)
+
+Disk pre-flight: 46 GB free on `/` (above the ~20 G build threshold) — no
+`target/` cleanup needed.
+
+```
+$ cargo test -p miroir-core --lib index_stats
+test result: ok. 9 passed; 0 failed
+
+$ cargo test -p miroir-core --lib compute_source_doc_count
+test result: ok. 4 passed; 0 failed
+```
+
+Whole `miroir-core` lib compiled clean (712 tests in the suite, no warnings).
+Both seams green: the pure reduction and the real-reqwest transport against
+mockito nodes.
+
+## Conclusion
+
+All four acceptance criteria are met and verified. The reshard backfill now has
+the missing denominator (source index real document count, computed before the
+first page is fetched) available via `compute_source_document_count`, computed
+identically to ILM. Umbrella bf-2ynu5 closes; the remaining work (populating the
+persisted `ReshardOperation.total_documents` / `start_backfill` wiring) is
+tracked under other children of bf-49a17.