Replace overflowing ASCII architecture with a native Mermaid block diagram

README.md

@@ -111,36 +111,15 @@ print(out.saliency_scores.shape)   # (1, 512)       per-token importance
 
 ## Architecture
 
-```
-RGB Image (B, 3, H, W)
-      │
-      ▼
-┌───────────────────────────┐
-│ CNN Stem (4-stage)         │  3→128→256→512→768, cumulative stride 16×
-└───────────────────────────┘
-      │
-      ├──────────────► Saliency Predictor  →  per-pixel importance map [0,1]
-      ▼
-┌───────────────────────────┐
-│ ATB Tokenizer              │  Gumbel-Top-K + NMS + differentiable weighted pooling
-│  budget split:             │  global 5% · region 25% · detail 70%
-│  512 default (up to 2048)  │
-└───────────────────────────┘
-      │
-      ▼
-┌───────────────────────────┐
-│ HCT levels                 │  global / region / detail + learned level embeddings
-└───────────────────────────┘
-      │
-      ▼
-┌───────────────────────────┐
-│ Spatial Transformer (24L)  │  RMSNorm · SwiGLU · continuous 2D-RoPE
-│  + [POOL] token            │  log-area scale bias (8 buckets)
-└───────────────────────────┘
-      │
-      ▼
-  pooled (B, 768)  +  tokens (B, N, 768)
-  + level_ids + spatial_metadata + saliency_scores
+```mermaid
+flowchart TD
+    IMG["RGB Image · (B,3,H,W)"] --> STEM["CNN Stem (4-stage)<br/>3 → 128 → 256 → 512 → 768 · stride 16×"]
+    STEM --> SAL["Saliency Predictor<br/>per-pixel importance map [0,1]"]
+    STEM --> ATB["ATB Tokenizer · Gumbel-Top-K + NMS<br/>weighted pooling · budget 512 (up to 2048)<br/>global 5% · region 25% · detail 70%"]
+    SAL --> ATB
+    ATB --> HCT["HCT levels<br/>global / region / detail + learned level embeddings"]
+    HCT --> TF["Spatial Transformer ×24<br/>RMSNorm · SwiGLU · continuous 2D-RoPE<br/>log-area scale bias (8 buckets) · [POOL] token"]
+    TF --> OUT["pooled (B,768) + tokens (B,N,768)<br/>+ level_ids + spatial_metadata + saliency_scores"]
 ```
 
 The `pooled` image embedding is the `[POOL]` token after the final RMSNorm (no separate projection head). For contrastive training it is aligned against a trainable projection of a frozen CLIP text encoder.