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.

5 files changed, 179 insertions, 115 deletions

diff --git a/cnn_v3/shaders/cnn_v3_bottleneck.wgsl b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
index e30682b..09819cc 100644
--- a/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
+++ b/cnn_v3/shaders/cnn_v3_bottleneck.wgsl
@@ -1,43 +1,49 @@
 // CNN v3 — Bottleneck
-// AvgPool2x2(enc1) + Conv(8->8, 3x3, dilation=2) + ReLU  (no FiLM)
+// AvgPool2x2(enc1) + Conv(16->16, 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_lo  (rgba32uint, 8xf16)  half-res  ch 0-7
+//         enc1_tex_hi  (rgba32uint, 8xf16)  half-res  ch 8-15
+// Output: bn_out_lo    (rgba32uint, 8xf16)  quarter-res
+//         bn_out_hi    (rgba32uint, 8xf16)  quarter-res
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 8*8*9)  conv: w[out][in][ky*3+kx]  (3x3 kernel, OIHW)
-//   [576 .. +8)   bias: b[out]
+//   [0 .. 16*16*9)  conv: w[out][in][ky*3+kx]
+//   [2304 .. +16)   bias: b[out]
 
 #include "cnn_v3/common"
 
-const BN_IN:      u32 = 8u;
-const BN_OUT:     u32 = 8u;
+const BN_IN:      u32 = 16u;
+const BN_OUT:     u32 = 16u;
 const BN_DILATION: i32 = 2;
 
 struct Params {
     weight_offset: u32,
-    _pad0: u32, _pad1: u32, _pad2: u32,  // 3 explicit pads: array<u32,3> invalid in uniform
+    _pad0: u32, _pad1: u32, _pad2: u32,
 }
 
-@group(0) @binding(0) var enc1_tex: texture_2d<u32>;
-@group(0) @binding(1) var<storage, read> weights: array<u32>;
-@group(0) @binding(2) var<uniform> params: Params;
-@group(0) @binding(3) var bottleneck_out: texture_storage_2d<rgba32uint, write>;
+@group(0) @binding(0) var enc1_tex_lo: texture_2d<u32>;
+@group(0) @binding(1) var enc1_tex_hi: 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 bn_out_lo: texture_storage_2d<rgba32uint, write>;
+@group(0) @binding(5) var bn_out_hi: 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 3x3 conv).
-fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array<f32, 8> {
+fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array<f32, 16> {
     let quart_dims = half_dims / 2;
     if (qcoord.x < 0 || qcoord.y < 0 || qcoord.x >= quart_dims.x || qcoord.y >= quart_dims.y) {
-        return array<f32, 8>(0., 0., 0., 0., 0., 0., 0., 0.);
+        return array<f32, 16>(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.);
     }
     let base = qcoord * 2;
     var s: array<f32, BN_IN>;
     for (var dy: i32 = 0; dy < 2; dy++) {
         for (var dx: i32 = 0; dx < 2; dx++) {
             let hc = clamp(base + vec2i(dx, dy), vec2i(0), half_dims - vec2i(1));
-            let f  = unpack_8ch(enc1_tex, hc);
-            for (var i: u32 = 0u; i < BN_IN; i++) { s[i] += f[i]; }
+            let lo = unpack_8ch(enc1_tex_lo, hc);
+            let hi = unpack_8ch(enc1_tex_hi, hc);
+            for (var i: u32 = 0u; i < 8u; i++) {
+                s[i]      += lo[i];
+                s[i + 8u] += hi[i];
+            }
         }
     }
     for (var i: u32 = 0u; i < BN_IN; i++) { s[i] *= 0.25; }
@@ -46,7 +52,7 @@ fn load_enc1_avg(qcoord: vec2i, half_dims: vec2i) -> array<f32, 8> {
 
 @compute @workgroup_size(8, 8)
 fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) {
-    let half_dims  = vec2i(textureDimensions(enc1_tex));
+    let half_dims  = vec2i(textureDimensions(enc1_tex_lo));
     let quart_dims = half_dims / 2;
     let coord      = vec2i(id.xy);
     if (coord.x >= quart_dims.x || coord.y >= quart_dims.y) { return; }
@@ -55,7 +61,7 @@ fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) {
     var out: array<f32, BN_OUT>;
 
     for (var o: u32 = 0u; o < BN_OUT; o++) {
-        var sum = get_w(wo, BN_OUT * BN_IN * 9u + o);  // bias (at end of 3x3 conv weights)
+        var sum = get_w(wo, BN_OUT * BN_IN * 9u + o);  // bias
         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);
@@ -68,10 +74,16 @@ fn bottleneck_main(@builtin(global_invocation_id) id: vec3u) {
         out[o] = max(0.0, sum);
     }
 
-    textureStore(bottleneck_out, coord, vec4u(
+    textureStore(bn_out_lo, 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]))
     ));
+    textureStore(bn_out_hi, coord, vec4u(
+        pack2x16float(vec2f(out[8],  out[9])),
+        pack2x16float(vec2f(out[10], out[11])),
+        pack2x16float(vec2f(out[12], out[13])),
+        pack2x16float(vec2f(out[14], out[15]))
+    ));
 }
diff --git a/cnn_v3/shaders/cnn_v3_dec0.wgsl b/cnn_v3/shaders/cnn_v3_dec0.wgsl
index a2a70ac..617b5a2 100644
--- a/cnn_v3/shaders/cnn_v3_dec0.wgsl
+++ b/cnn_v3/shaders/cnn_v3_dec0.wgsl
@@ -1,19 +1,17 @@
 // CNN v3 — Decoder level 0 + output
-// NearestUp2x(dec1) + cat(enc0_skip) -> Conv(8->4, 3x3, zero-pad) + FiLM + ReLU + Sigmoid
+// NearestUp2x(dec1) + cat(enc0_skip) -> Conv(16->4, 3x3) + FiLM + ReLU + Sigmoid
 //
-// Inputs: dec1_tex  (rgba16float, 4ch)  half-res
-//         enc0_tex  (rgba16float, 4ch)  full-res  (skip connection)
-// Output: output_tex (rgba16float, 4ch) full-res  (dispatch at full-res dims)
+// Inputs: dec1_tex   (rgba32uint, 8xf16)  half-res
+//         enc0_tex   (rgba32uint, 8xf16)  full-res  (skip connection)
+// Output: output_tex (rgba16float, 4ch)   full-res
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 8*4*9)  conv: w[out][in][ky][kx]   (in=8: 4 dec1 + 4 enc0 skip)
-//   [288 .. +4)   bias: b[out]
-//
-// Parity note: sigmoid applied after FiLM+ReLU, not after raw conv (matches train_cnn_v3.py).
+//   [0 .. 16*4*9)  conv: w[out][in][ky][kx]  (in=16: 8 dec1 + 8 enc0 skip)
+//   [576 .. +4)    bias: b[out]
 
 #include "cnn_v3/common"
 
-const DEC0_IN:  u32 = 8u;
+const DEC0_IN:  u32 = 16u;
 const DEC0_OUT: u32 = 4u;
 
 struct Params {
@@ -23,25 +21,27 @@ struct Params {
     beta:  vec4f,
 }
 
-@group(0) @binding(0) var dec1_tex: texture_2d<f32>;
-@group(0) @binding(1) var enc0_tex: texture_2d<f32>;
+@group(0) @binding(0) var dec1_tex: texture_2d<u32>;
+@group(0) @binding(1) var enc0_tex: 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 output_tex: texture_storage_2d<rgba16float, write>;
 
-// Load 8 concatenated channels at full-res coord:
-//   ch 0-3:  dec1 nearest-up (dec1_tex[coord/2])
-//   ch 4-7:  enc0 skip       (enc0_tex[coord])
-// Returns zeros for OOB coord (zero-padding for the conv).
-fn load_dec0_concat(coord: vec2i, full_dims: vec2i) -> array<f32, 8> {
+// Load 16ch: ch 0-7 from dec1 nearest-up, ch 8-15 from enc0 skip.
+fn load_dec0_concat(coord: vec2i, full_dims: vec2i) -> array<f32, 16> {
+    var r: array<f32, 16>;
     if (coord.x < 0 || coord.y < 0 || coord.x >= full_dims.x || coord.y >= full_dims.y) {
-        return array<f32, 8>(0., 0., 0., 0., 0., 0., 0., 0.);
+        return r;
     }
     let half_dims = vec2i(textureDimensions(dec1_tex));
-    let hc  = clamp(coord / 2, vec2i(0), half_dims - vec2i(1));
-    let d   = textureLoad(dec1_tex, hc, 0);
-    let e   = textureLoad(enc0_tex, coord, 0);
-    return array<f32, 8>(d.x, d.y, d.z, d.w, e.x, e.y, e.z, e.w);
+    let hc = clamp(coord / 2, vec2i(0), half_dims - vec2i(1));
+    let d  = unpack_8ch(dec1_tex, hc);
+    let e  = unpack_8ch(enc0_tex, coord);
+    for (var i: u32 = 0u; i < 8u; i++) {
+        r[i]      = d[i];
+        r[i + 8u] = e[i];
+    }
+    return r;
 }
 
 @compute @workgroup_size(8, 8)
@@ -64,7 +64,6 @@ fn dec0_main(@builtin(global_invocation_id) id: vec3u) {
                 }
             }
         }
-        // FiLM + ReLU + Sigmoid (matches training forward())
         let v = max(0.0, params.gamma[o] * sum + params.beta[o]);
         out[o] = 1.0 / (1.0 + exp(-v));
     }
diff --git a/cnn_v3/shaders/cnn_v3_dec1.wgsl b/cnn_v3/shaders/cnn_v3_dec1.wgsl
index 28ae3dc..fadea3b 100644
--- a/cnn_v3/shaders/cnn_v3_dec1.wgsl
+++ b/cnn_v3/shaders/cnn_v3_dec1.wgsl
@@ -1,53 +1,71 @@
 // CNN v3 — Decoder level 1
-// NearestUp2x(bottleneck) + cat(enc1_skip) -> Conv(16->4, 3x3, zero-pad) + FiLM + ReLU
+// NearestUp2x(bottleneck) + cat(enc1_skip) -> Conv(32->8, 3x3) + FiLM + ReLU
 //
-// Inputs: bottleneck_tex (rgba32uint, 8xf16)  quarter-res
-//         enc1_tex       (rgba32uint, 8xf16)  half-res    (skip connection)
-// Output: dec1_out       (rgba16float, 4ch)   half-res    (dispatch at half-res dims)
+// Inputs: bn_tex_lo   (rgba32uint, 8xf16)  quarter-res  ch 0-7
+//         bn_tex_hi   (rgba32uint, 8xf16)  quarter-res  ch 8-15
+//         enc1_tex_lo (rgba32uint, 8xf16)  half-res     skip ch 0-7
+//         enc1_tex_hi (rgba32uint, 8xf16)  half-res     skip ch 8-15
+// Output: dec1_out    (rgba32uint, 8xf16)  half-res
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 16*4*9)  conv: w[out][in][ky][kx]   (in=16: 8 bottleneck + 8 enc1 skip)
-//   [576 .. +4)    bias: b[out]
+//   [0 .. 32*8*9)   conv: w[out][in][ky][kx]  (in=32: 16 bn + 16 enc1 skip)
+//   [2304 .. +8)    bias: b[out]
 
 #include "cnn_v3/common"
 
-const DEC1_IN:  u32 = 16u;
-const DEC1_OUT: u32 = 4u;
+const DEC1_IN:  u32 = 32u;
+const DEC1_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 bottleneck_tex: texture_2d<u32>;
-@group(0) @binding(1) var enc1_tex: 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 dec1_out: texture_storage_2d<rgba16float, write>;
+@group(0) @binding(0) var bn_tex_lo:   texture_2d<u32>;
+@group(0) @binding(1) var bn_tex_hi:   texture_2d<u32>;
+@group(0) @binding(2) var enc1_tex_lo: texture_2d<u32>;
+@group(0) @binding(3) var enc1_tex_hi: texture_2d<u32>;
+@group(0) @binding(4) var<storage, read> weights: array<u32>;
+@group(0) @binding(5) var<uniform> params: Params;
+@group(0) @binding(6) var dec1_out: texture_storage_2d<rgba32uint, write>;
 
-// Load 16 concatenated channels at half-res coord hcoord:
-//   ch 0-7:  bottleneck nearest-up (bottleneck_tex[hcoord/2])
-//   ch 8-15: enc1 skip             (enc1_tex[hcoord])
-// Returns zeros for OOB hcoord (zero-padding for the conv).
-fn load_dec1_concat(hcoord: vec2i, half_dims: vec2i) -> array<f32, 16> {
+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];
+}
+
+// Load 32ch: [bn_nearest_up(16ch), enc1_skip(16ch)]
+fn load_dec1_concat(hcoord: vec2i, half_dims: vec2i) -> array<f32, 32> {
+    var r: array<f32, 32>;
     if (hcoord.x < 0 || hcoord.y < 0 || hcoord.x >= half_dims.x || hcoord.y >= half_dims.y) {
-        return array<f32, 16>(0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.,0.);
+        return r;
     }
     let quart_dims = half_dims / 2;
-    let qc = clamp(hcoord / 2, vec2i(0), quart_dims - vec2i(1));
-    let b  = unpack_8ch(bottleneck_tex, qc);
-    let s  = unpack_8ch(enc1_tex, hcoord);
-    return array<f32, 16>(
-        b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7],
-        s[0], s[1], s[2], s[3], s[4], s[5], s[6], s[7]
-    );
+    let qc  = clamp(hcoord / 2, vec2i(0), quart_dims - vec2i(1));
+    let blo = unpack_8ch(bn_tex_lo,   qc);
+    let bhi = unpack_8ch(bn_tex_hi,   qc);
+    let slo = unpack_8ch(enc1_tex_lo, hcoord);
+    let shi = unpack_8ch(enc1_tex_hi, hcoord);
+    for (var i: u32 = 0u; i < 8u; i++) {
+        r[i]       = blo[i];
+        r[i + 8u]  = bhi[i];
+        r[i + 16u] = slo[i];
+        r[i + 24u] = shi[i];
+    }
+    return r;
 }
 
 @compute @workgroup_size(8, 8)
 fn dec1_main(@builtin(global_invocation_id) id: vec3u) {
-    let half_dims = vec2i(textureDimensions(enc1_tex));
+    let half_dims = vec2i(textureDimensions(enc1_tex_lo));
     let coord     = vec2i(id.xy);
     if (coord.x >= half_dims.x || coord.y >= half_dims.y) { return; }
 
@@ -65,8 +83,13 @@ fn dec1_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(dec1_out, coord, vec4f(out[0], out[1], out[2], out[3]));
+    textureStore(dec1_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]))
+    ));
 }
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]))
+    ));
 }
diff --git a/cnn_v3/shaders/cnn_v3_enc1.wgsl b/cnn_v3/shaders/cnn_v3_enc1.wgsl
index 23e485d..eb41279 100644
--- a/cnn_v3/shaders/cnn_v3_enc1.wgsl
+++ b/cnn_v3/shaders/cnn_v3_enc1.wgsl
@@ -1,58 +1,67 @@
 // CNN v3 — Encoder level 1
-// AvgPool2x2(enc0) + Conv(4->8, 3x3, zero-pad) + FiLM + ReLU
+// AvgPool2x2(enc0) + Conv(8->16, 3x3, zero-pad) + FiLM + ReLU
 //
-// Input:  enc0_tex (rgba16float, 4ch)   full-res
-// Output: enc1_out (rgba32uint, 8xf16)  half-res  (dispatch at half-res dims)
+// Input:  enc0_tex     (rgba32uint, 8xf16)  full-res
+// Output: enc1_out_lo  (rgba32uint, 8xf16)  half-res  ch 0-7
+//         enc1_out_hi  (rgba32uint, 8xf16)  half-res  ch 8-15
 //
 // Weight layout (f16, OIHW + bias):
-//   [0 .. 4*8*9)  conv: w[out][in][ky][kx]
-//   [288 .. +8)   bias: b[out]
+//   [0 .. 8*16*9)   conv: w[out][in][ky][kx]
+//   [1152 .. +16)   bias: b[out]
 
 #include "cnn_v3/common"
 
-const ENC1_IN:  u32 = 4u;
-const ENC1_OUT: u32 = 8u;
+const ENC1_IN:  u32 = 8u;
+const ENC1_OUT: u32 = 16u;
 
 struct Params {
     weight_offset: u32,
     _pad: vec3u,
-    gamma_lo: vec4f,   // FiLM gamma ch 0-3
-    gamma_hi: vec4f,   // FiLM gamma ch 4-7
-    beta_lo:  vec4f,   // FiLM beta  ch 0-3
-    beta_hi:  vec4f,   // FiLM beta  ch 4-7
+    gamma_0: vec4f,
+    gamma_1: vec4f,
+    gamma_2: vec4f,
+    gamma_3: vec4f,
+    beta_0:  vec4f,
+    beta_1:  vec4f,
+    beta_2:  vec4f,
+    beta_3:  vec4f,
 }
 
-@group(0) @binding(0) var enc0_tex: texture_2d<f32>;
+@group(0) @binding(0) var enc0_tex: texture_2d<u32>;
 @group(0) @binding(1) var<storage, read> weights: array<u32>;
 @group(0) @binding(2) var<uniform> params: Params;
-@group(0) @binding(3) var enc1_out: texture_storage_2d<rgba32uint, write>;
+@group(0) @binding(3) var enc1_out_lo: texture_storage_2d<rgba32uint, write>;
+@group(0) @binding(4) var enc1_out_hi: 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];
+    if (o < 4u)  { return params.gamma_0[o]; }
+    if (o < 8u)  { return params.gamma_1[o - 4u]; }
+    if (o < 12u) { return params.gamma_2[o - 8u]; }
+    return params.gamma_3[o - 12u];
 }
 fn film_beta(o: u32) -> f32 {
-    if (o < 4u) { return params.beta_lo[o]; }
-    return params.beta_hi[o - 4u];
+    if (o < 4u)  { return params.beta_0[o]; }
+    if (o < 8u)  { return params.beta_1[o - 4u]; }
+    if (o < 12u) { return params.beta_2[o - 8u]; }
+    return params.beta_3[o - 12u];
 }
 
-// Avg-pool 2x2 from enc0_tex at half-res coord hcoord.
-// Returns zeros for OOB half-res coords (zero-padding for the conv).
-fn load_enc0_avg(hcoord: vec2i, full_dims: vec2i) -> array<f32, 4> {
+fn load_enc0_avg(hcoord: vec2i, full_dims: vec2i) -> array<f32, 8> {
     let half_dims = full_dims / 2;
     if (hcoord.x < 0 || hcoord.y < 0 || hcoord.x >= half_dims.x || hcoord.y >= half_dims.y) {
-        return array<f32, 4>(0., 0., 0., 0.);
+        return array<f32, 8>(0., 0., 0., 0., 0., 0., 0., 0.);
     }
     let base = hcoord * 2;
-    var s = vec4f(0.);
+    var s: array<f32, 8>;
     for (var dy: i32 = 0; dy < 2; dy++) {
         for (var dx: i32 = 0; dx < 2; dx++) {
             let fc = clamp(base + vec2i(dx, dy), vec2i(0), full_dims - vec2i(1));
-            s += textureLoad(enc0_tex, fc, 0);
+            let f  = unpack_8ch(enc0_tex, fc);
+            for (var i: u32 = 0u; i < 8u; i++) { s[i] += f[i]; }
         }
     }
-    let avg = s * 0.25;
-    return array<f32, 4>(avg.x, avg.y, avg.z, avg.w);
+    for (var i: u32 = 0u; i < 8u; i++) { s[i] *= 0.25; }
+    return s;
 }
 
 @compute @workgroup_size(8, 8)
@@ -79,10 +88,16 @@ fn enc1_main(@builtin(global_invocation_id) id: vec3u) {
         out[o] = max(0.0, film_gamma(o) * sum + film_beta(o));
     }
 
-    textureStore(enc1_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]))
+    textureStore(enc1_out_lo, 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]))
+    ));
+    textureStore(enc1_out_hi, coord, vec4u(
+        pack2x16float(vec2f(out[8],  out[9])),
+        pack2x16float(vec2f(out[10], out[11])),
+        pack2x16float(vec2f(out[12], out[13])),
+        pack2x16float(vec2f(out[14], out[15]))
     ));
 }