Scope the speed figure to the CLS-only build that produced it

README.md

@@ -95,11 +95,13 @@ accuracy across datasets. The supplied package retains its original conversion m
 
 On an M4 Pro running macOS 26.3, native optimized Swift serial inference achieved
 26.6 images/second across 400 images, with median latency 37.39 ms and p95 39.68 ms.
-The benchmark used Core ML compute units `.all`, five warmups, and three passes;
-throughput uses the median pass duration. Input feature creation, prediction, and
-vector extraction were timed. Photo fetching, resizing, model loading, and database
-writes were excluded. These are local measurements, not a hardware-independent guarantee.
-Only aggregate timing results are included here.
+This was measured on an earlier build that returned only `embedding`; reading
+`patch_embeddings` as well copies a further 2.4 MB per image, which that figure does
+not include. The benchmark used Core ML compute units `.all`, five warmups, and three
+passes; throughput uses the median pass duration. Input feature creation, prediction,
+and vector extraction were timed. Photo fetching, resizing, model loading, and
+database writes were excluded. These are local measurements, not a hardware-independent
+guarantee. Only aggregate timing results are included here.
 
 ## Attribution and license