feat(graph): T99.1.2 mark Gemma4PLECombinedProducer non-capturable

graph/cuda_graph.go

@@ -44,14 +44,22 @@ var debugGraphCapture = os.Getenv("ZERFOO_DEBUG_GPU") == "1"
 // (offset_memcpy kernel) and GQA uses GPU RoPE selection (rope_select kernel),
 // all position-dependent state is read from GPU memory at replay time, making
 // GQA fully capturable.
+//
+// Gemma4PLECombinedProducer: performs a CPU-side gather over the shared PLE
+// embedding table (token ids -> per-layer rows), then calls MulScalar on the
+// freshly-allocated CPUStorage tensor. Running this inside a capture stream
+// triggers a synchronous H2D cudaMemcpy that CUDA rejects. The producer runs
+// once per forward pass before the transformer loop, so placing it in
+// pre-capture keeps the layer-body capture region intact. See ADR-088.
 var nonCapturableOps = map[string]bool{
-	"EmbeddingLookup":   true,
-	"Gather":            true,
-	"AutoAttentionMask": true,
-	"AutoPositionIds":   true,
-	"Slice":             true,
-	"ConstantOfShape":   true,
-	"Shape":             true,
+	"EmbeddingLookup":            true,
+	"Gather":                     true,
+	"AutoAttentionMask":          true,
+	"AutoPositionIds":            true,
+	"Slice":                      true,
+	"ConstantOfShape":            true,
+	"Shape":                      true,
+	"Gemma4PLECombinedProducer":  true,
 }
 
 // isNonCapturable returns true if the instruction at index i in the plan