Fix download progress bugs: false completion, stale UI, dedup

Three fixes on top of the download progress feature:

1. Backend: Replace broken "no .incomplete = done" completion check
   with a 95% byte threshold. HF downloads files sequentially, so
   between files there are briefly no .incomplete files even though
   the download is far from done (e.g. Kimi-K2.5 reported "done"
   after downloading 22KB of config files out of 595GB).

2. Frontend: Track hasShownProgress flag. Only show "Download
   complete. Loading into memory..." if we actually displayed
   download progress before. For already-cached models where the
   first poll returns progress=1.0, this avoids the misleading
   "Download complete" message.

3. Frontend: Deduplicate recommended vs downloaded -- filter out
   models already in the "Downloaded" section. Cache the fetched
   lists at module level so re-mounting the popover does not flash
   an empty "Downloaded" section.

studio/backend/routes/models.py

@@ -738,9 +738,12 @@ async def get_download_progress(
                 "progress": 0,
             }
 
-        # No .incomplete files means download is done regardless of byte totals
-        # (blob deduplication can make completed_bytes differ from expected_bytes)
-        if in_progress_bytes == 0 and completed_bytes > 0:
+        # Use 95% threshold for completion (blob deduplication can make
+        # completed_bytes differ slightly from expected_bytes).
+        # Do NOT use "no .incomplete files" as a completion signal --
+        # HF downloads files sequentially, so between files there are
+        # no .incomplete files even though the download is far from done.
+        if completed_bytes >= expected_bytes * 0.95:
             progress = 1.0
         else:
             progress = min(downloaded_bytes / expected_bytes, 0.99)

studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx

@@ -340,6 +340,10 @@ function isGgufRepo(id: string): boolean {
   return id.toUpperCase().includes("-GGUF");
 }
 
+// Module-level caches so re-mounting the popover shows results instantly
+let _cachedGgufCache: CachedGgufRepo[] = [];
+let _cachedModelsCache: CachedModelRepo[] = [];
+
 // ── Hub Model Picker ──────────────────────────────────────────
 
 export function HubModelPicker({
@@ -361,17 +365,27 @@ export function HubModelPicker({
   // Track which GGUF repo is expanded for variant selection
   const [expandedGguf, setExpandedGguf] = useState<string | null>(null);
 
-  // Cached (already downloaded) repos
-  const [cachedGguf, setCachedGguf] = useState<CachedGgufRepo[]>([]);
-  const [cachedModels, setCachedModels] = useState<CachedModelRepo[]>([]);
+  // Cached (already downloaded) repos -- use module-level cache so
+  // re-mounting the popover does not flash an empty "Downloaded" section.
+  const [cachedGguf, setCachedGguf] = useState<CachedGgufRepo[]>(_cachedGgufCache);
+  const [cachedModels, setCachedModels] = useState<CachedModelRepo[]>(_cachedModelsCache);
   useEffect(() => {
-    listCachedGguf().then(setCachedGguf).catch(() => {});
-    listCachedModels().then(setCachedModels).catch(() => {});
+    listCachedGguf().then((v) => { _cachedGgufCache = v; setCachedGguf(v); }).catch(() => {});
+    listCachedModels().then((v) => { _cachedModelsCache = v; setCachedModels(v); }).catch(() => {});
   }, []);
 
+  // Deduplicate: don't show downloaded models in the recommended list
+  const downloadedSet = useMemo(() => {
+    const s = new Set<string>();
+    for (const c of cachedGguf) s.add(c.repo_id);
+    for (const c of cachedModels) s.add(c.repo_id);
+    return s;
+  }, [cachedGguf, cachedModels]);
+
   const recommendedIds = useMemo(
-    () => dedupe([...models.map((model) => model.id), value ?? ""]),
-    [models, value],
+    () => dedupe([...models.map((model) => model.id), value ?? ""])
+      .filter((id) => !downloadedSet.has(id)),
+    [models, value, downloadedSet],
   );
 
   const { paramCountById: recommendedParamCountById } =

studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts

@@ -347,6 +347,8 @@ export function useChatModelRuntime() {
             },
           };
 
+          let hasShownProgress = false;
+
           const pollProgress = async () => {
             // Stop if cancelled or if loading already finished
             if (abortCtrl.signal.aborted || !loadingModelRef.current) {
@@ -361,7 +363,8 @@ export function useChatModelRuntime() {
               // Re-check after await -- load may have finished while polling
               if (!loadingModelRef.current) return;
 
-              if (prog.downloaded_bytes > 0 && prog.progress > 0 && prog.progress < 1) {
+              if (prog.progress > 0 && prog.progress < 1) {
+                hasShownProgress = true;
                 const dlGb = prog.downloaded_bytes / (1024 ** 3);
                 const totalGb = prog.expected_bytes / (1024 ** 3);
                 const pct = Math.round(prog.progress * 100);
@@ -376,7 +379,9 @@ export function useChatModelRuntime() {
                     action: cancelAction,
                   },
                 );
-              } else if (prog.downloaded_bytes > 0 && prog.expected_bytes === 0) {
+              } else if (prog.downloaded_bytes > 0 && prog.expected_bytes === 0 && prog.progress === 0) {
+                // Have bytes but no total size -- show bytes only
+                hasShownProgress = true;
                 const dlGb = prog.downloaded_bytes / (1024 ** 3);
                 toast.loading(
                   "Downloading model...",
@@ -387,7 +392,8 @@ export function useChatModelRuntime() {
                     action: cancelAction,
                   },
                 );
-              } else if (prog.progress >= 1) {
+              } else if (prog.progress >= 1 && hasShownProgress) {
+                // Only show "download complete" if we actually showed progress
                 toast.loading("Loading model...", {
                   id: toastId,
                   description: "Download complete. Loading into memory...",