fix: render video via native <video> HTML to bypass gradio SSRF-download of HF xet CDN

app.py

@@ -139,11 +139,20 @@ def _format_meta(model: str, task: dict[str, Any], idx: int, total: int) -> str:
 
 def _load_item(pool_order: list[int], idx: int) -> tuple[str, str, str]:
     if idx < 0 or idx >= len(pool_order):
-        return "", "**All done!** No more items.", ""
+        return "<div style='padding:24px;color:#888'>All done — no more items.</div>", "**All done!** No more items.", ""
     model, task_id = POOL[pool_order[idx]]
     task = TASK_BY_ID[task_id]
+    url = _video_url(model, task_id)
+    # Render a native <video> so the browser streams directly from HF CDN,
+    # bypassing gradio's server-side SSRF-protected download (which blocks
+    # the xet-bridge hostnames).
+    video_html = (
+        f'<video src="{url}" controls autoplay loop muted playsinline '
+        f'style="width:100%;max-height:520px;background:#000;border-radius:8px;">'
+        f'Your browser does not support HTML5 video.</video>'
+    )
     return (
-        _video_url(model, task_id),
+        video_html,
         _format_meta(model, task, idx, len(pool_order)),
         _extract_prompt(task),
     )
@@ -161,9 +170,9 @@ def start_session(annotator: str, state: dict):
     rng = random.Random(f"{annotator}-{int(time.time())}")
     rng.shuffle(order)
     state = {"annotator": annotator, "order": order, "idx": 0}
-    video, meta, prompt = _load_item(order, 0)
+    video_html, meta, prompt = _load_item(order, 0)
     status = f"✅ Logged in as `{annotator}` — {len(order)} items to annotate."
-    return state, video, meta, prompt, status
+    return state, video_html, meta, prompt, status
 
 
 def submit_and_next(state: dict, score: float, note: str):
@@ -186,16 +195,16 @@ def submit_and_next(state: dict, score: float, note: str):
     }
     _append_annotation(record)
     state["idx"] = idx + 1
-    video, meta, prompt = _load_item(state["order"], state["idx"])
-    return state, video, meta, prompt, 3, "", f"✅ Submitted ({state['idx']}). Next →"
+    video_html, meta, prompt = _load_item(state["order"], state["idx"])
+    return state, video_html, meta, prompt, 3, "", f"✅ Submitted ({state['idx']}). Next →"
 
 
 def skip_and_next(state: dict):
     if not state or "order" not in state:
         return state, "", "⚠️ Please log in first.", "", 3, "", "⚠️ Not logged in."
     state["idx"] = state["idx"] + 1
-    video, meta, prompt = _load_item(state["order"], state["idx"])
-    return state, video, meta, prompt, 3, "", f"⏭️ Skipped. Position: {state['idx']}"
+    video_html, meta, prompt = _load_item(state["order"], state["idx"])
+    return state, video_html, meta, prompt, 3, "", f"⏭️ Skipped. Position: {state['idx']}"
 
 
 # ---------------------------------------------------------------------------
@@ -228,7 +237,12 @@ with gr.Blocks(title="MBench-V Annotation", theme=gr.themes.Soft()) as demo:
 
     with gr.Row():
         with gr.Column(scale=3):
-            video = gr.Video(label="Generated video", autoplay=True, loop=True)
+            video = gr.HTML(
+                value="<div style='padding:24px;color:#888;text-align:center'>"
+                      "Enter your name above and click <b>Start</b> to load the first video."
+                      "</div>",
+                label="Generated video",
+            )
         with gr.Column(scale=2):
             meta_md = gr.Markdown()
             prompt_tb = gr.Textbox(