MONA only on last 12 blocks (skip early low-level blocks)

- Add last_n_blocks parameter to inject_mona_into_dinov3 (default=12)
- Blocks 0-11: pure frozen DINOv3 (low-level features, domain-agnostic)
- Blocks 12-23: MONA adapted (high-level semantic features)
- MONA params: 3.5M (was 6.85M, -49%)
- Total trainable: ~5.7M with text (was ~9.0M)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/models/adapters.py

@@ -149,24 +149,32 @@ def inject_mona_into_dinov3(
     model: nn.Module,
     bottleneck: int = 64,
     dropout: float = 0.1,
+    last_n_blocks: int = 12,
 ) -> int:
     """Inject MONA adapters into a frozen DINOv3ViT model.
 
     Adds two MonaAdapter modules per block (after attention, after MLP).
-    Only adapter parameters are trainable.
+    Only adapter parameters are trainable. Early blocks remain pure frozen
+    DINOv3 (low-level features don't need spatial adaptation).
 
     Args:
         model: DINOv3ViT model (already frozen).
         bottleneck: MONA bottleneck dimension.
         dropout: MONA dropout rate.
+        last_n_blocks: Number of last blocks to adapt (default 12 of 24).
 
     Returns:
         Number of trainable adapter parameters added.
     """
     dim = model.embed_dim
     n_adapters = 0
+    total_blocks = len(model.layer)
+    start_idx = total_blocks - last_n_blocks
+
+    for i, block in enumerate(model.layer):
+        if i < start_idx:
+            continue  # Skip early blocks — frozen, no adaptation.
 
-    for block in model.layer:
         # Create adapters.
         block.mona_attn = MonaAdapter(dim, bottleneck, dropout)
         block.mona_mlp = MonaAdapter(dim, bottleneck, dropout)
@@ -195,8 +203,8 @@ def inject_mona_into_dinov3(
 
     n_params = sum(p.numel() for n, p in model.named_parameters() if "mona" in n)
     LOGGER.info(
-        "🔧 MONA injected: %d adapters (%d per block), %s trainable params (bottleneck=%d)",
-        n_adapters, 2, f"{n_params:,}", bottleneck,
+        "🔧 MONA injected: %d adapters (blocks %d-%d of %d), %s trainable params (bottleneck=%d)",
+        n_adapters, start_idx, total_blocks - 1, total_blocks, f"{n_params:,}", bottleneck,
     )
     return n_params