feat(cnn_v3): upgrade architecture to enc_channels=[8,16]

Double encoder capacity: enc0 4→8ch, enc1 8→16ch, bottleneck 16→16ch,
dec1 32→8ch, dec0 16→4ch. Total weights 2476→7828 f16 (~15.3 KB).
FiLM MLP output 40→72 params (L1: 16×40→16×72).

16-ch textures split into _lo/_hi rgba32uint pairs (enc1, bottleneck).
enc0 and dec1 textures changed from rgba16float to rgba32uint (8ch).
GBUF_RGBA32UINT node gains CopySrc for parity test readback.

- WGSL shaders: all 5 passes rewritten for new channel counts
- C++ CNNv3Effect: new weight offsets/sizes, 8ch uniform structs
- Web tool (shaders.js + tester.js): matching texture formats and bindings
- Parity test: readback_rgba32uint_8ch helper, updated vector counts
- Training scripts: default enc_channels=[8,16], updated docstrings
- Docs + architecture PNG regenerated

handoff(Gemini): CNN v3 [8,16] upgrade complete. All code, tests, web
tool, training scripts, and docs updated. Next: run training pass.

1 files changed, 26 insertions, 11 deletions

diff --git a/cnn_v3/shaders/cnn_v3_enc0.wgsl b/cnn_v3/shaders/cnn_v3_enc0.wgsl
index e171ca7..84d40fd 100644
--- a/cnn_v3/shaders/cnn_v3_enc0.wgsl
+++ b/cnn_v3/shaders/cnn_v3_enc0.wgsl
@@ -1,32 +1,42 @@
 // CNN v3 — Encoder level 0
-// Conv(20->4, 3x3, zero-pad) + FiLM + ReLU
+// Conv(20->8, 3x3, zero-pad) + FiLM + ReLU
 //
 // Input:  feat_tex0 (rgba32uint, 8xf16), feat_tex1 (rgba32uint, 12ch u8norm)  full-res
-// Output: enc0_out  (rgba16float, 4ch)                                   full-res
+// Output: enc0_out  (rgba32uint, 8xf16)                                         full-res
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 20*4*9)  conv: w[out][in][ky][kx]
-//   [720 .. +4)    bias: b[out]
+//   [0 .. 20*8*9)  conv: w[out][in][ky][kx]
+//   [1440 .. +8)   bias: b[out]
 
 #include "cnn_v3/common"
 
 const ENC0_IN:  u32 = 20u;
-const ENC0_OUT: u32 = 4u;
+const ENC0_OUT: u32 = 8u;
 
 struct Params {
     weight_offset: u32,
     _pad: vec3u,
-    gamma: vec4f,
-    beta:  vec4f,
+    gamma_lo: vec4f,
+    gamma_hi: vec4f,
+    beta_lo:  vec4f,
+    beta_hi:  vec4f,
 }
 
 @group(0) @binding(0) var feat_tex0: texture_2d<u32>;
 @group(0) @binding(1) var feat_tex1: texture_2d<u32>;
 @group(0) @binding(2) var<storage, read> weights: array<u32>;
 @group(0) @binding(3) var<uniform> params: Params;
-@group(0) @binding(4) var enc0_out: texture_storage_2d<rgba16float, write>;
+@group(0) @binding(4) var enc0_out: texture_storage_2d<rgba32uint, write>;
+
+fn film_gamma(o: u32) -> f32 {
+    if (o < 4u) { return params.gamma_lo[o]; }
+    return params.gamma_hi[o - 4u];
+}
+fn film_beta(o: u32) -> f32 {
+    if (o < 4u) { return params.beta_lo[o]; }
+    return params.beta_hi[o - 4u];
+}
 
-// Unpack all 20 feature channels at coord. Returns zeros for OOB (zero-padding).
 fn load_feat(coord: vec2i, dims: vec2i) -> array<f32, 20> {
     if (coord.x < 0 || coord.y < 0 || coord.x >= dims.x || coord.y >= dims.y) {
         return array<f32, 20>(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.);
@@ -68,8 +78,13 @@ fn enc0_main(@builtin(global_invocation_id) id: vec3u) {
                 }
             }
         }
-        out[o] = max(0.0, params.gamma[o] * sum + params.beta[o]);
+        out[o] = max(0.0, film_gamma(o) * sum + film_beta(o));
     }
 
-    textureStore(enc0_out, coord, vec4f(out[0], out[1], out[2], out[3]));
+    textureStore(enc0_out, coord, vec4u(
+        pack2x16float(vec2f(out[0], out[1])),
+        pack2x16float(vec2f(out[2], out[3])),
+        pack2x16float(vec2f(out[4], out[5])),
+        pack2x16float(vec2f(out[6], out[7]))
+    ));
 }