fix(cnn_v3): fix texture format mismatches in cnn_v3_test sequence

- seq_compiler: add gbuf_albedo/gbuf_rgba32uint to NODE_TYPES
- timeline: declare gbuf_feat0/feat1 as gbuf_rgba32uint, route
  CNNv3Effect output through cnn_v3_out (gbuf_albedo) + Passthrough
  to sink (dec0 can't write directly to Rgba8Unorm sink)
- cnn_v3_effect: fix update_bind_groups using .set() instead of
  .replace() causing FATAL assert on second frame
- TODO: add CNN v3 "2D mode" (G-buffer-free) future task

handoff(Gemini): CNNv3Effect now runs without crashes at --seek 48

4 files changed, 29 insertions, 6 deletions

diff --git a/TODO.md b/TODO.md
index 48518f2..4f38740 100644
--- a/TODO.md
+++ b/TODO.md
@@ -84,6 +84,23 @@ PyTorch / HTML WebGPU / C++ WebGPU.
 - ✅ `export_cnn_v3_weights.py` — convert trained `.pth` → `.bin` (f16)
 - See `cnn_v3/docs/HOWTO.md` §3 for training commands
 
+## Future: CNN v3 "2D Mode" (G-buffer-free)
+
+Allow `CNNv3Effect` to run on a plain screen buffer / photo without a real G-buffer.
+Fake the missing feature vectors (normals, depth, material IDs, shadow, transp) from
+the RGB input alone:
+- normals: approximate from local luminance gradient (Sobel)
+- depth: constant (e.g. 0.5) or estimated from a simple heuristic
+- material IDs / shadow / transp: neutral defaults (e.g. 0)
+
+This would let the effect be applied to any rendered frame (post-NTSC, post-Scratch, etc.)
+without requiring a 3D G-buffer pass upstream, and enable training/inference on photos.
+
+Implementation sketch:
+- New `CNNv3Effect2D` subclass (or a mode flag) that synthesizes `feat_tex0`/`feat_tex1`
+  internally from a single `rgba8unorm` input, then runs the same 5-pass U-Net.
+- Separate `gbuf_pack_2d.wgsl` compute shader that fills feat0/feat1 from a photo buffer.
+
 ## Future: CNN v2 8-bit Quantization
 
 Reduce weights from f16 (~3.2 KB) to i8 (~1.6 KB).
diff --git a/cnn_v3/src/cnn_v3_effect.cc b/cnn_v3/src/cnn_v3_effect.cc
index 4aa2c25..bfbb17b 100644
--- a/cnn_v3/src/cnn_v3_effect.cc
+++ b/cnn_v3/src/cnn_v3_effect.cc
@@ -445,7 +445,7 @@ void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
     bg_buf(e[2], 2, wb, kWeightsBufBytes);
     bg_buf(e[3], 3, enc0_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
     bg_tex(e[4], 4, enc0_view);
-    enc0_bg_.set(make_bg(enc0_pipeline_.get(), e, 5));
+    enc0_bg_.replace(make_bg(enc0_pipeline_.get(), e, 5));
   }
 
   // enc1: enc0_tex(B0), weights(B1), params(B2), enc1_out(B3)
@@ -455,7 +455,7 @@ void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
     bg_buf(e[1], 1, wb, kWeightsBufBytes);
     bg_buf(e[2], 2, enc1_params_buf_.get().buffer, sizeof(CnnV3ParamsEnc1));
     bg_tex(e[3], 3, enc1_view);
-    enc1_bg_.set(make_bg(enc1_pipeline_.get(), e, 4));
+    enc1_bg_.replace(make_bg(enc1_pipeline_.get(), e, 4));
   }
 
   // bottleneck: enc1_tex(B0), weights(B1), params(B2), bn_out(B3)
@@ -465,7 +465,7 @@ void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
     bg_buf(e[1], 1, wb, kWeightsBufBytes);
     bg_buf(e[2], 2, bn_params_buf_.get().buffer, sizeof(CnnV3ParamsBn));
     bg_tex(e[3], 3, bn_view);
-    bn_bg_.set(make_bg(bn_pipeline_.get(), e, 4));
+    bn_bg_.replace(make_bg(bn_pipeline_.get(), e, 4));
   }
 
   // dec1: bn_tex(B0), enc1_tex(B1), weights(B2), params(B3), dec1_out(B4)
@@ -476,7 +476,7 @@ void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
     bg_buf(e[2], 2, wb, kWeightsBufBytes);
     bg_buf(e[3], 3, dec1_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
     bg_tex(e[4], 4, dec1_view);
-    dec1_bg_.set(make_bg(dec1_pipeline_.get(), e, 5));
+    dec1_bg_.replace(make_bg(dec1_pipeline_.get(), e, 5));
   }
 
   // dec0: dec1_tex(B0), enc0_tex(B1), weights(B2), params(B3), output(B4)
@@ -487,6 +487,6 @@ void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
     bg_buf(e[2], 2, wb, kWeightsBufBytes);
     bg_buf(e[3], 3, dec0_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
     bg_tex(e[4], 4, out_view);
-    dec0_bg_.set(make_bg(dec0_pipeline_.get(), e, 5));
+    dec0_bg_.replace(make_bg(dec0_pipeline_.get(), e, 5));
   }
 }
diff --git a/tools/seq_compiler.py b/tools/seq_compiler.py
index fbd5c0d..09188a5 100755
--- a/tools/seq_compiler.py
+++ b/tools/seq_compiler.py
@@ -18,6 +18,8 @@ NODE_TYPES = {
     'f16x8': 'NodeType::F16X8',
     'depth24': 'NodeType::DEPTH24',
     'compute_f32': 'NodeType::COMPUTE_F32',
+    'gbuf_albedo': 'NodeType::GBUF_ALBEDO',
+    'gbuf_rgba32uint': 'NodeType::GBUF_RGBA32UINT',
 }
 
 class NodeDecl:
diff --git a/workspaces/main/timeline.seq b/workspaces/main/timeline.seq
index 1a9cad3..2176275 100644
--- a/workspaces/main/timeline.seq
+++ b/workspaces/main/timeline.seq
@@ -44,5 +44,9 @@ SEQUENCE 40.00 0 "ntsc"
   EFFECT + Ntsc temp1 -> sink 0.00 8.00
 
 SEQUENCE 48.00 0 "cnn_v3_test"
+  NODE gbuf_feat0 gbuf_rgba32uint
+  NODE gbuf_feat1 gbuf_rgba32uint
+  NODE cnn_v3_out gbuf_albedo
   EFFECT + GBufferEffect source -> gbuf_feat0 gbuf_feat1 0.00 8.00
-  EFFECT + CNNv3Effect gbuf_feat0 gbuf_feat1 -> sink 0.00 8.00
+  EFFECT + CNNv3Effect gbuf_feat0 gbuf_feat1 -> cnn_v3_out 0.00 8.00
+  EFFECT + Passthrough cnn_v3_out -> sink 0.00 8.00