CNN v2: Fix activation function mismatch between training and inference

Layer 0 now uses clamp [0,1] in both training and inference (was using ReLU in shaders).
- index.html: Add is_layer_0 flag to LayerParams, handle Layer 0 separately
- export_cnn_v2_shader.py: Generate correct activation for Layer 0

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

1 files changed, 5 insertions, 2 deletions

diff --git a/training/export_cnn_v2_shader.py b/training/export_cnn_v2_shader.py
index ad5749c..dc475d8 100755
--- a/training/export_cnn_v2_shader.py
+++ b/training/export_cnn_v2_shader.py
@@ -35,9 +35,12 @@ def export_layer_shader(layer_idx, weights, kernel_size, output_dir, is_output_l
     )
 
     radius = kernel_size // 2
-    activation = "" if is_output_layer else "output[c] = max(0.0, sum);  // ReLU"
     if is_output_layer:
-        activation = "output[c] = clamp(sum, 0.0, 1.0);  // Sigmoid approximation"
+        activation = "output[c] = clamp(sum, 0.0, 1.0);  // Output layer"
+    elif layer_idx == 0:
+        activation = "output[c] = clamp(sum, 0.0, 1.0);  // Layer 0: clamp [0,1]"
+    else:
+        activation = "output[c] = max(0.0, sum);  // Middle layers: ReLU"
 
     shader_code = f"""// CNN v2 Layer {layer_idx} - Auto-generated (uniform 12D→4D)
 // Kernel: {kernel_size}×{kernel_size}, In: 12D (4 prev + 8 static), Out: 4D