feat(cnn_v3): 3×3 dilated bottleneck + Sobel loss + FiLM warmup + architecture PNG

- Replace 1×1 pointwise bottleneck with Conv(8→8, 3×3, dilation=2):
  effective RF grows from ~13px to ~29px at ¼res (~+1 KB weights)
- Add Sobel edge loss in training (--edge-loss-weight, default 0.1)
- Add FiLM 2-phase training: freeze MLP for warmup epochs then
  unfreeze at lr×0.1 (--film-warmup-epochs, default 50)
- Update weight layout: BN 72→584 f16, total 1964→2476 f16 (4952 B)
- Cascade offsets in C++ effect, JS tool, export/gen_test_vectors scripts
- Regenerate test_vectors.h (1238 u32); parity max_err=9.77e-04
- Generate dark-theme U-Net+FiLM architecture PNG (gen_architecture_png.py)
- Replace ASCII art in CNN_V3.md and HOW_TO_CNN.md with PNG embed

handoff(Gemini): bottleneck dilation + Sobel loss + FiLM warmup landed.
Next: run first real training pass (see cnn_v3/docs/HOWTO.md §3).

1 files changed, 19 insertions, 13 deletions

diff --git a/cnn_v3/shaders/cnn_v3_bottleneck.wgsl b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
index e24586f..e30682b 100644
--- a/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
+++ b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
@@ -1,17 +1,18 @@
 // CNN v3 — Bottleneck
-// AvgPool2x2(enc1) + Conv(8->8, 1x1) + ReLU  (no FiLM)
+// AvgPool2x2(enc1) + Conv(8->8, 3x3, dilation=2) + ReLU  (no FiLM)
 //
-// Input:  enc1_tex      (rgba32uint, 8xf16)  half-res
-// Output: bottleneck_out (rgba32uint, 8xf16) quarter-res  (dispatch at quarter-res dims)
+// Input:  enc1_tex       (rgba32uint, 8xf16)  half-res
+// Output: bottleneck_out (rgba32uint, 8xf16)  quarter-res  (dispatch at quarter-res dims)
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 8*8*1)  conv: w[out][in]  (1x1 kernel)
-//   [64 .. +8)    bias: b[out]
+//   [0 .. 8*8*9)  conv: w[out][in][ky*3+kx]  (3x3 kernel, OIHW)
+//   [576 .. +8)   bias: b[out]
 
 #include "cnn_v3/common"
 
-const BN_IN:  u32 = 8u;
-const BN_OUT: u32 = 8u;
+const BN_IN:      u32 = 8u;
+const BN_OUT:     u32 = 8u;
+const BN_DILATION: i32 = 2;
 
 struct Params {
     weight_offset: u32,
@@ -24,7 +25,7 @@ struct Params {
 @group(0) @binding(3) var bottleneck_out: texture_storage_2d<rgba32uint, write>;
 
 // Avg-pool 2x2 from enc1_tex at quarter-res coord qcoord.
-// Returns zeros for OOB quarter-res coords (zero-padding for the 1x1 conv).
+// Returns zeros for OOB quarter-res coords (zero-padding for the 3x3 conv).
 fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array<f32, 8> {
     let quart_dims = half_dims / 2;
     if (qcoord.x < 0 || qcoord.y < 0 || qcoord.x >= quart_dims.x || qcoord.y >= quart_dims.y) {
@@ -50,14 +51,19 @@ fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) {
     let coord      = vec2i(id.xy);
     if (coord.x >= quart_dims.x || coord.y >= quart_dims.y) { return; }
 
-    let wo   = params.weight_offset;
-    let feat = load_enc1_avg(coord, half_dims);
+    let wo  = params.weight_offset;
     var out: array<f32, BN_OUT>;
 
     for (var o: u32 = 0u; o < BN_OUT; o++) {
-        var sum = get_w(wo, BN_OUT * BN_IN + o);  // bias (1x1 kernel: no spatial idx)
-        for (var i: u32 = 0u; i < BN_IN; i++) {
-            sum += get_w(wo, o * BN_IN + i) * feat[i];
+        var sum = get_w(wo, BN_OUT * BN_IN * 9u + o);  // bias (at end of 3x3 conv weights)
+        for (var ky: i32 = -1; ky <= 1; ky++) {
+            for (var kx: i32 = -1; kx <= 1; kx++) {
+                let feat = load_enc1_avg(coord + vec2i(kx, ky) * BN_DILATION, half_dims);
+                let ki   = u32(ky + 1) * 3u + u32(kx + 1);
+                for (var i: u32 = 0u; i < BN_IN; i++) {
+                    sum += get_w(wo, o * BN_IN * 9u + i * 9u + ki) * feat[i];
+                }
+            }
         }
         out[o] = max(0.0, sum);
     }