feat(settings): DirectML GPU acceleration package card (P2)

Mirrors the NPU package card, shown only when hardware_status marks DirectML
applicable; downloads model+runtime via download_directml_package and reloads
hardware to flip to Ready.

src/lib/views/Settings.svelte

@@ -269,6 +269,25 @@
     }
   }
 
+  // ---- DirectML package download (P2) ----
+  // ponytail: the await resolves when model+runtime are both staged, so the
+  // busy flag is the whole progress story here — no % listener like NPU has.
+  let dmlBusy = $state(false);
+  let dmlMsg = $state<string | null>(null);
+  async function downloadDirectmlPackage() {
+    dmlBusy = true;
+    dmlMsg = "Downloading…";
+    try {
+      await api.downloadDirectmlPackage();
+      await settings.loadHardware();
+      dmlMsg = "DirectML package ready.";
+    } catch (e) {
+      dmlMsg = `Failed: ${errorMessage(e)}`;
+    } finally {
+      dmlBusy = false;
+    }
+  }
+
   onMount(() => {
     settings.load();
     calendar.load();
@@ -486,6 +505,35 @@
               {#if npuMsg}<p class="muted npu-status">{npuMsg}</p>{/if}
             </div>
           {/if}
+
+          {#if settings.hardware.directml?.applicable}
+            {@const dml = settings.hardware.directml}
+            {@const dmlReady = dml.runtimeReady && dml.modelInstalled}
+            <!-- ponytail: reuses the .npu-package layout styles (shared accel card). -->
+            <div class="npu-package">
+              <div class="row">
+                <Cpu size={14} aria-hidden="true" />
+                <strong>GPU acceleration (DirectML)</strong>
+                {#if dmlReady}
+                  <span class="badge ok"><Check size={12} aria-hidden="true" /> Ready</span>
+                {:else if dmlBusy}
+                  <span class="badge busy">Downloading…</span>
+                {:else}
+                  <span class="badge">Package needed</span>
+                {/if}
+              </div>
+              {#if !dmlReady}
+                <p class="muted">
+                  Your GPU can be accelerated without Vulkan via DirectML — a one-time package
+                  (Whisper ONNX model{dml.runtimeReady ? "" : " + DirectML runtime"}).
+                </p>
+                <button onclick={downloadDirectmlPackage} disabled={dmlBusy}>
+                  {dmlBusy ? "Downloading…" : "Download DirectML package"}
+                </button>
+              {/if}
+              {#if dmlMsg}<p class="muted npu-status">{dmlMsg}</p>{/if}
+            </div>
+          {/if}
         {:else}
           <p class="muted">Hardware detection unavailable.</p>
         {/if}