cuda: clean up failed and unwound Laya graph captures

.github/workflows/ci.yml

@@ -23,13 +23,13 @@ jobs:
       - name: Format
         run: cargo fmt --all --check
       - name: Clippy
-        run: cargo clippy --workspace --locked --all-targets -- -D warnings
+        run: cargo clippy --workspace --locked --all-targets --features omni-laya/serve -- -D warnings
       - name: Frontend API tests
         run: cargo test -p omni-jev --test frontend --locked
       - name: Test
-        run: cargo test --workspace --locked
+        run: cargo test --workspace --locked --features omni-laya/serve
       - name: Build
-        run: cargo build --workspace --release --locked
+        run: cargo build --workspace --release --locked --features omni-laya/serve
 
   benchmarks:
     runs-on: ubuntu-latest

CONTRIBUTING.md

@@ -76,9 +76,9 @@ The Rust checks used by [CI](.github/workflows/ci.yml) are:
 
 ```sh
 cargo fmt --all --check
-cargo clippy --workspace --locked --all-targets -- -D warnings
-cargo test --workspace --locked
-cargo build --workspace --release --locked
+cargo clippy --workspace --locked --all-targets --features omni-laya/serve -- -D warnings
+cargo test --workspace --locked --features omni-laya/serve
+cargo build --workspace --release --locked --features omni-laya/serve
 ```
 
 For an agent-assisted self-review, use the repository's
@@ -96,29 +96,6 @@ remain responsible for understanding the changes and verifying the results.
 Self-review helps maintainers focus on design and correctness; it does not
 replace maintainer review.
 
-### Large code changes
-
-PRs with **more than 3,000 changed lines of authored code** need extra contributor
-attention before requesting review. Count additions plus deletions against the
-PR's merge base in source files, tests, and build or validation scripts. Report
-this count separately from the total diff size; exclude documentation, generated
-output, lockfiles, and static fixtures from the code count, while still reviewing
-those files for relevance and correctness.
-
-- Complete a full self-review of every affected component and its integration
-  boundaries. A quick precheck alone is insufficient; keep the PR in draft until
-  the contributor self-review is complete.
-- Consider splitting independent features, refactors, and cleanup into focused
-  PRs. If the change needs to stay together, explain why in the PR description
-  and provide a component map and suggested review order.
-- Include the code-line count and a validation summary for each affected area in
-  the PR description: commands, results, and unverified behavior with reasons.
-  Cover changed interfaces between components as well as individual components.
-
-Size signals the need for closer review; it is not itself a correctness finding.
-Choose checks based on the changed behavior and risk. Crossing this threshold
-alone does not require GPU benchmarks or other expensive experiments.
-
 ## Documentation site
 
 The site at <https://thinkflowlab.github.io/system1-omni/> is built with MkDocs

src/backends/cuda/Cargo.toml

@@ -7,3 +7,10 @@ publish = false
 [dependencies]
 anyhow = "1"
 libloading = "0.8"
+
+[dev-dependencies]
+tempfile = "3"
+
+[[test]]
+name = "laya_runtime"
+path = "../../../tests/cuda/laya_runtime.rs"

src/backends/cuda/kernels/runtime.cu

@@ -62,14 +62,26 @@ int laya_capture_begin(void* stream) {
                                 cudaStreamCaptureModeThreadLocal);
 }
 
+// End capture and release temporary/partial graph ownership on every outcome.
 int laya_capture_end(void* stream, void** executable) {
+  if (!executable) return -1;
+  *executable = nullptr;
+  if (!stream) return -1;
   cudaGraph_t graph = nullptr;
-  auto e = cudaStreamEndCapture(static_cast<cudaStream_t>(stream), &graph);
-  if (e != cudaSuccess)
-    return e;
-  e = cudaGraphInstantiate(reinterpret_cast<cudaGraphExec_t*>(executable), graph, 0);
-  cudaGraphDestroy(graph);
-  return e;
+  cudaGraphExec_t created = nullptr;
+  auto status = cudaStreamEndCapture(static_cast<cudaStream_t>(stream), &graph);
+  if (status == cudaSuccess && graph)
+    status = cudaGraphInstantiate(&created, graph, 0);
+  if (graph) {
+    auto destroyed = cudaGraphDestroy(graph);
+    if (status == cudaSuccess) status = destroyed;
+  }
+  if (status != cudaSuccess || !created) {
+    if (created) cudaGraphExecDestroy(created);
+    return status != cudaSuccess ? status : -1;
+  }
+  *executable = created;
+  return 0;
 }
 
 int laya_graph_run(void* executable, void* stream) {