gpu, gpu-gemma-embedding2, gpu-gates: the EmbeddingGemma 2 encoder from the common kernels, with its end-to-end gate

Cargo.toml

@@ -1,6 +1,6 @@
 [workspace]
 resolver = "3"
-members = ["crates/qdot", "crates/gguf", "crates/models", "crates/model", "crates/placement", "crates/threads", "crates/levers", "crates/verbs", "crates/gpu", "crates/gpu-deepseek41", "crates/gpu-glm5next", "crates/gpu-mimo2", "crates/gpu-gates", "crates/engram", "crates/engram-lab", "crates/sampler", "crates/tokenizer", "crates/serve", "crates/jinja", "crates/vision", "crates/gpu-vision", "crates/refset", "crates/runtime", "crates/app", "crates/decision", "crates/hf"]
+members = ["crates/qdot", "crates/gguf", "crates/models", "crates/model", "crates/placement", "crates/threads", "crates/levers", "crates/verbs", "crates/gpu", "crates/gpu-deepseek41", "crates/gpu-glm5next", "crates/gpu-mimo2", "crates/gpu-gemma-embedding2", "crates/gpu-gates", "crates/engram", "crates/engram-lab", "crates/sampler", "crates/tokenizer", "crates/serve", "crates/jinja", "crates/vision", "crates/gpu-vision", "crates/refset", "crates/runtime", "crates/app", "crates/decision", "crates/hf"]
 # 의도적으로 컴파일되지 않는 재현기다(MUL-7). 워크스페이스에 넣으면 --workspace가 영구히 빨갛다.
 exclude = ["crates/oxide-ice-unroll", "crates/oxide-ice-unroll/variants/unroll_smoke", "fuzz", "labs/hip-native"]
 

crates/gpu-gates/Cargo.toml

@@ -44,6 +44,9 @@ bloomery-gpu-deepseek41 = { path = "../gpu-deepseek41", optional = true }
 bloomery-gpu-glm5next = { path = "../gpu-glm5next", optional = true }
 # The MiMo-V2.6-Flash program, reached only through the `mimo2` feature (its end-to-end gate).
 bloomery-gpu-mimo2 = { path = "../gpu-mimo2", optional = true }
+# The EmbeddingGemma 2 encoder, reached only through the `gemma-embedding2` feature (its end-to-end
+# gate).
+bloomery-gpu-gemma-embedding2 = { path = "../gpu-gemma-embedding2", optional = true }
 cuda-core = { workspace = true, optional = true }
 # A gate binary's own `#[cuda_module]` (`gate_kquant`: entries no engine binary links) needs the
 # device and host halves of cuda-oxide as direct dependencies: the macros emit `::cuda_host` and
@@ -119,6 +122,7 @@ vision = ["gpu", "dep:bloomery-gpu-vision", "dep:bloomery-vision"]
 clef = ["gpu", "dep:bloomery-decision", "dep:bloomery-serve", "dep:serde_json"]
 glm5next = ["deepseek41", "dep:bloomery-gpu-glm5next", "bloomery-app/glm5next"]
 mimo2 = ["gpu", "dep:bloomery-gpu-mimo2", "bloomery-app/mimo2"]
+gemma-embedding2 = ["gpu", "dep:bloomery-gpu-gemma-embedding2"]
 
 # A bin whose `main` sits behind a feature requires that feature: cargo refuses to build it without the
 # feature before it links anything, instead of linking the feature-off stub `main` over the working
@@ -535,6 +539,10 @@ required-features = ["gpu"]
 name = "gate_gemma_embedding2_head512"
 required-features = ["gpu"]
 
+[[bin]]
+name = "gate_gemma_embedding2_e2e"
+required-features = ["gemma-embedding2"]
+
 [[bin]]
 name = "gate_qwen3moe_router"
 required-features = ["gpu"]