card: device-verified numbers (iPhone 17 Pro re-prompt 1.3 s; tracker per-frame)

README.md

@@ -28,8 +28,9 @@ match's box, score, and 288×288 instance mask. No server, no cloud.
 - **ViT-L/14 trunk @1008²** + tri-neck → **CLIP-L text encoder** → **text-conditioned DETR
   head** (200 queries, presence token) — ~830 M params.
 - Verified on a **Pixel 8a** (vision GPU 9.2 s / text CPU 0.5 s / head GPU 1.4 s; re-prompt
-  1.9 s; kept-set equal to PyTorch fp32, mask IoU ≥ 0.98) and on **Apple Metal** (exact with
-  `enforce_f32`).
+  1.9 s; kept-set equal to PyTorch fp32, mask IoU ≥ 0.98) and on an **iPhone 17 Pro**
+  (first prompt ≈ 5.7 s; **re-prompt on the same photo ≈ 1.3 s** — vision features are
+  cached per image, so changing the phrase is nearly instant).
 - Every graph re-authoring is **exact** (corr 1.0 vs PyTorch) — no approximations.
 
 ![SAM 3 on-device: "car"](assets/street_car.png)
@@ -112,6 +113,9 @@ keep = np.where(prob > 0.5)[0]                                     # boxes y[200
 recondition, memory bank + temporal pos-enc) is ported and verified against the official
 model (ids identical, mask IoU ≥ 0.992) — the executable spec and the Kotlin/Swift ports
 live in the **LiteRT-Models** zoo sample (`sam3/`), together with all conversion scripts.
+A 48-frame clip tracks end-to-end on an iPhone 17 Pro at ≈16.6 s/frame (vision 7.3 s,
+memory attention 6.4 s, head 2.7 s on CPU, decoders <0.2 s) and at ≈2 s/frame on an
+M4 Max — offline processing rather than real time.
 
 ## Conversion notes