@@ -74,12 +74,24 @@ const harness = Effect.gen(function* () {
       );
     }),
   );
-  const store = yield* SpeechModelStore.SpeechModelStore.pipe(
+  const openStore = SpeechModelStore.SpeechModelStore.pipe(
     Effect.provide(
       SpeechModelStore.layer.pipe(Layer.provide(Layer.succeed(HttpClient.HttpClient, client))),
     ),
   );
-  return { model, directory, stale, fs, requests, bodies, transportFailure, bodyStarted, store };
+  const store = yield* openStore;
+  return {
+    model,
+    directory,
+    stale,
+    fs,
+    requests,
+    bodies,
+    transportFailure,
+    bodyStarted,
+    store,
+    openStore,
+  };
 });
 
 const dependencies = ServerConfig.layerTest("/tmp", { prefix: "speech-model-store-" }).pipe(
@@ -136,7 +148,7 @@ it.effect("rejects a same-size download whose SHA-256 differs from the catalog",
   }).pipe(Effect.provide(dependencies)),
 );
 
-it.effect("cancelling a download removes the partial file and keeps the existing model", () =>
+it.effect("cleans only download partials on cancellation or restart", () =>
   Effect.gen(function* () {
     const h = yield* harness;
     h.bodies.current = () => Stream.concat(Stream.succeed(bytes.subarray(0, SPLIT)), Stream.never);
@@ -145,5 +157,13 @@ it.effect("cancelling a download removes the partial file and keeps the existing
     yield* Fiber.interrupt(fiber);
     expect(yield* h.fs.readDirectory(h.directory)).toEqual(["model.gguf"]);
     expect([...(yield* h.fs.readFile(h.stale))]).toEqual([...new Uint8Array(bytes.length)]);
+    const orphan = `${h.stale}.f660de6c-e172-4146-8ffe-d3c32eb48cd4.part`;
+    yield* h.fs.writeFile(orphan, bytes.subarray(0, SPLIT));
+    yield* h.fs.writeFileString(`${h.stale}.notes.part`, "not a download");
+    yield* h.openStore;
+    expect((yield* h.fs.readDirectory(h.directory)).sort()).toEqual([
+      "model.gguf",
+      "model.gguf.notes.part",
+    ]);
   }).pipe(Effect.provide(dependencies)),
 );
@@ -54,6 +54,8 @@ export class SpeechModelStore extends Context.Service<
 >()("t3/speech/SpeechModelStore") {}
 
 const DOWNLOAD_ATTEMPTS = 3;
+const PARTIAL_FILENAME =
+  /\.gguf\.[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}\.part$/i;
 
 const make = Effect.gen(function* () {
   const fs = yield* FileSystem.FileSystem;
@@ -62,6 +64,12 @@ const make = Effect.gen(function* () {
   const random = yield* SecureRandom.SecureRandom;
   const config = yield* ServerConfig.ServerConfig;
   const directory = path.join(config.stateDir, "speech", "models");
+  // The server already owns this state directory; no download can be active before the store opens.
+  for (const name of yield* fs.readDirectory(directory).pipe(Effect.orElseSucceed(() => []))) {
+    if (PARTIAL_FILENAME.test(name)) {
+      yield* fs.remove(path.join(directory, name), { force: true }).pipe(Effect.ignore);
+    }
+  }
   const modelPath = (model: SpeechModel) => path.join(directory, model.filename);
   // Files whose digest failed, by identity, so status polls stop calling them ready.
   const corrupt = new Map<string, string>();
 
@@ -12,6 +12,7 @@ import * as Exit from "effect/Exit";
 import * as HostProcess from "@t3tools/shared/HostProcess";
 import { DEFAULT_SPEECH_TRANSCRIPTION_OPTIONS } from "@t3tools/contracts";
 import * as ServerSettings from "../serverSettings.ts";
+import { listNativeSpeechGpuDevices } from "./native.ts";
 import * as SpeechModelStore from "./SpeechModelStore.ts";
 import * as SpeechService from "./SpeechService.ts";
 
@@ -398,6 +399,24 @@ it.effect("reports unsupported hosts without wrapping a synthetic cause", () =>
   }),
 );
 
+it.effect.each([{ devices: [] }, { devices: [{ id: "gpu", name: "Test GPU" }] }])(
+  "retries failed GPU discovery without repeatedly probing on status polls: $devices",
+  ({ devices }) =>
+    Effect.gen(function* () {
+      const speech = yield* SpeechService.SpeechService;
+      const discover = vi.mocked(listNativeSpeechGpuDevices).mockReset();
+      discover.mockRejectedValueOnce(new Error("driver not ready")).mockResolvedValueOnce(devices);
+      expect(yield* speech.status).toMatchObject({ gpuDevices: [] });
+      yield* speech.status;
+      expect(discover).toHaveBeenCalledOnce();
+      yield* TestClock.adjust("1 minute");
+      expect(yield* speech.status).toMatchObject({ gpuDevices: devices });
+      yield* TestClock.adjust("1 minute");
+      expect(yield* speech.status).toMatchObject({ gpuDevices: devices });
+      expect(discover).toHaveBeenCalledTimes(2);
+    }).pipe(Effect.provide(layer)),
+);
+
 it.effect(
   "retries Auto-placed inference with the dictionary and the CPU's vocabulary capability",
   () =>
 
@@ -10,6 +10,7 @@ import * as HostProcess from "@t3tools/shared/HostProcess";
 import * as Cause from "effect/Cause";
 import * as Clock from "effect/Clock";
 import * as Context from "effect/Context";
+import * as Duration from "effect/Duration";
 import * as Effect from "effect/Effect";
 import * as Exit from "effect/Exit";
 import * as Fiber from "effect/Fiber";
@@ -183,7 +184,9 @@ const make = Effect.gen(function* () {
   let loadedAcceleration: string | undefined;
   let autoCpuModel: string | undefined;
   let loading: Promise<LoadedModel> | undefined;
-  let gpuDevices: Promise<{ id: string; name: string }[]> | undefined;
+  const discoverGpuDevices = yield* Effect.tryPromise(() => listNativeSpeechGpuDevices()).pipe(
+    Effect.cachedWithTTL((exit) => (Exit.isSuccess(exit) ? Duration.infinity : "1 minute")),
+  );
   let downloading:
     | {
         modelId: string;
@@ -438,9 +441,7 @@ const make = Effect.gen(function* () {
       supportsLanguageDetection: definition.supportsLanguageDetection,
       acceleration: settings.speechAcceleration,
       modelUnloadTimeout: settings.speechModelUnloadTimeout,
-      gpuDevices: yield* Effect.promise(
-        () => (gpuDevices ??= listNativeSpeechGpuDevices().catch(() => [])),
-      ),
+      gpuDevices: yield* discoverGpuDevices.pipe(Effect.orElseSucceed(() => [])),
     };
   });