fix(inference): reject wrong-platform NEMOCLAW_VLLM_MODEL overrides before downloads

src/lib/inference/vllm-models.ts

@@ -74,8 +74,8 @@ export interface VllmModelDef {
    * platform-specific flags (the NVFP4 MoE checkpoint targets `sm_121a` only,
    * the very large V4 Flash recipe wants Station-class VRAM) appear only on
    * profiles they can actually run on. Direct `NEMOCLAW_VLLM_MODEL`
-   * overrides normally bypass the picker filter, but a model-specific runtime
-   * is rejected outside this list so an incompatible image is never pulled.
+   * overrides bypass the picker filter, so `runVllmInstall` rejects any
+   * override outside this list before the image pull and model download.
    */
   platforms: readonly VllmPlatform[];
   /**

src/lib/inference/vllm.test.ts

@@ -760,6 +760,66 @@ describe("installVllm model resolution", () => {
     );
   });
 
+  it("rejects a Spark-only override without a runtime before side effects on generic Linux (#7358)", async () => {
+    process.env.NEMOCLAW_VLLM_MODEL = "qwen3.6-35b-a3b-nvfp4";
+    const profile = detectVllmProfile({ platform: "linux", type: "nvidia" })!;
+    const beforeInstall = vi.fn();
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn(),
+      beforeInstall,
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(beforeInstall).not.toHaveBeenCalled();
+    expect(mocks.runCapture).not.toHaveBeenCalled();
+    expect(mocks.dockerPullWithProgressWatchdog).not.toHaveBeenCalled();
+    expect(mocks.dockerSpawn).not.toHaveBeenCalled();
+    const errors = errSpy.mock.calls.map((c: unknown[]) => String(c[0])).join("\n");
+    expect(errors).toContain("Qwen3.6 35B-A3B NVFP4 is not supported on Linux + NVIDIA GPU");
+  });
+
+  it("rejects the Station-only V4 Flash override before its 352 GB download on generic Linux (#7358)", async () => {
+    process.env.NEMOCLAW_VLLM_MODEL = "deepseek-v4-flash";
+    const profile = detectVllmProfile({ platform: "linux", type: "nvidia" })!;
+    const beforeInstall = vi.fn();
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn(),
+      beforeInstall,
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(beforeInstall).not.toHaveBeenCalled();
+    expect(mocks.dockerPullWithProgressWatchdog).not.toHaveBeenCalled();
+    expect(mocks.dockerSpawn).not.toHaveBeenCalled();
+    const errors = errSpy.mock.calls.map((c: unknown[]) => String(c[0])).join("\n");
+    expect(errors).toContain("DeepSeek V4 Flash is not supported on Linux + NVIDIA GPU");
+  });
+
+  it("still accepts a platform-matched override on its own platform (#7358)", async () => {
+    process.env.NEMOCLAW_VLLM_MODEL = "qwen3.6-35b-a3b-nvfp4";
+    const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
+    const promptFn = vi.fn<(q: string) => Promise<string>>();
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn,
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(promptFn).not.toHaveBeenCalled();
+    const summary = logSpy.mock.calls.map((c: unknown[]) => String(c[0])).join("\n");
+    expect(summary).toContain("Model: nvidia/Qwen3.6-35B-A3B-NVFP4 (NEMOCLAW_VLLM_MODEL override)");
+    const errors = errSpy.mock.calls.map((c: unknown[]) => String(c[0])).join("\n");
+    expect(errors).not.toContain("is not supported on");
+  });
+
   it("installs the complete Nemotron Ultra Station recipe without another selection", async () => {
     process.env.NEMOCLAW_VLLM_MODEL = "nemotron-3-ultra-550b-a55b";
     mocks.getGpuIndicesByName.mockReturnValue([0]);

src/lib/inference/vllm.ts

@@ -1328,7 +1328,13 @@ async function runVllmInstall(
   });
   if (!resolved) return { ok: false };
   const { model, source: modelSource } = resolved;
-  if (model.runtime && !model.platforms.includes(profile.platform)) {
+  // Platform-restricted models are filtered out of the interactive picker,
+  // but a direct NEMOCLAW_VLLM_MODEL override bypasses that filter, so this
+  // gate is the only platform enforcement on the env-override path. It must
+  // reject every wrong-platform model, not just runtime-carrying ones — an
+  // NVFP4 Spark checkpoint or a 352 GB Station recipe cannot serve here and
+  // must fail before the image pull and download (#7358).
+  if (!model.platforms.includes(profile.platform)) {
     console.error(`  vLLM install failed: ${model.label} is not supported on ${profile.name}`);
     return { ok: false };
   }