refactor(llama-cpp): clarify terminal cleanup contract

extensions/llama-cpp/src/inference-provider.test.ts

@@ -937,7 +937,7 @@ describe("llama.cpp inference provider", () => {
     expect(mocks.llamaDispose).toHaveBeenCalledOnce();
   });
 
-  it("keeps a failed best-effort cleanup as one disposal attempt", async () => {
+  it("keeps a failed cleanup terminal", async () => {
     await collectTestEvents();
     mocks.contextDispose.mockRejectedValueOnce(new Error("context cleanup failed"));
 

extensions/llama-cpp/src/inference-provider.ts

@@ -307,8 +307,9 @@ function disposeLlamaCppInferenceRuntime(state: LlamaCppInferenceRuntimeState):
     return state.disposePromise;
   }
   state.lifecycle = "closing";
-  // Plugin services stop once, and node-llama-cpp disposers mark themselves
-  // disposed before awaiting cleanup. The first disposal attempt is authoritative.
+  // node-llama-cpp disposers are one-shot and child cleanup releases the
+  // parent's disposal guard. Do not force parent cleanup after a child rejects:
+  // the retained guard can make that parent disposer wait forever.
   state.disposePromise = serialize(state, async () => {
     await disposeLoadedModel(state);
     if (state.llamaInstance) {