docs: update status header — note 10% bit-identical Q4 + retraction fix

ober

8b087908e941d439dd20c4324814b0cef1461f4e

diff --git a/jerboa-db.md b/jerboa-db.md
index 49a68c5..e18e70b 100644
--- a/jerboa-db.md
+++ b/jerboa-db.md
@@ -7,16 +7,18 @@ with a Datalog query engine and immutable time-travel over all data.
 **Status:** 2026-05-03 — **Core Datomic parity: complete.  MBrainz
 performance parity: shipped.** Core test suite is 37/37 passing
 (`make test`).  Five-phase performance plan landed between
-commits `439b9a7..7a45249`: pure-aggregate pushdown, DuckDB
+commits `439b9a7..b5a32ab`: pure-aggregate pushdown, DuckDB
 analytical fallback (with batched VALUES sync, `:in` substitution,
 auto-routing via `current-analytics-engine`), and the load-time
 fix that took 1% MBrainz ingest from ~47s to ~1s (a fulltext
 indexing O(n²) bug, not the originally-suspected RB-tree
 allocation pressure).  At 1% scale the 8-query suite finishes in
 under 3 seconds total; Q4 lands in 281 ms via DuckDB and Q8 in 9
-ms — both faster than published Datahike numbers.  See
-[MBrainz Benchmark Results](#mbrainz-benchmark-results) for the
-table.
+ms — both faster than published Datahike numbers.  At 10% scale
+Q4 stays bit-for-bit identical between Datalog and DuckDB at
+4,270,990 rows after a retraction-aware sync fix (`fbc3b6c`).
+See [MBrainz Benchmark Results](#mbrainz-benchmark-results) for
+the table.
 
 Implementation lives in `lib/jerboa-db/` (Jerboa library files using `#!chezscheme`
 + `(library ...)` form).  Built from scratch — no `(std mvcc)`, `(std datalog)`, or