fix(ui): disable broken tabs + add @spaces.GPU + MiniCPM trust_remote_code note

Main Space (app.py):
- Comment out 5 broken tabs: Nemotron, Voice, Image, OCR, Translation
- Document reasons: event loop error, unavailable backends, config issues
- Keep working tabs: Ask, Chat, Mesh, Marketplace, Files, Emergency, Settings
- Add MiniCPM3-4B trust_remote_code requirement note to ask.py docstring

Nemotron Space (app_nemotron.py):
- Add 'import spaces' with graceful fallback
- Decorate extract_structured() with @spaces.GPU (or no-op if HAS_SPACES=False)
- Fixes 'No @spaces.GPU function detected during startup' error on HF ZeroGPU

app_nemotron.py

@@ -26,6 +26,13 @@ import os
 
 import gradio as gr
 
+# HF Spaces GPU support
+try:
+    import spaces
+    HAS_SPACES = True
+except ImportError:
+    HAS_SPACES = False
+
 # ── Optional mesh connection ──────────────────────────────────────────────────
 _MESH_NODE = os.getenv("HEARTHNET_NODE", "")
 _NVIDIA_KEY = os.getenv("NVIDIA_API_KEY", "")
@@ -148,6 +155,7 @@ async def _nemotron_chat(messages: list, model: str, api_key: str, temperature:
         return r.json()["choices"][0]["message"]["content"]
 
 
+@spaces.GPU if HAS_SPACES else lambda f: f
 def extract_structured(
     doc_text: str,
     schema_preset: str,
@@ -155,6 +163,11 @@ def extract_structured(
     model_label: str,
     api_key: str,
 ) -> tuple[str, str]:
+    """Extract structured data from documents using Nemotron.
+    
+    Wrapped with @spaces.GPU to signal GPU usage to HF Spaces.
+    Falls back gracefully if GPU unavailable (e.g., local testing).
+    """
     import json
 
     if not doc_text.strip():

hearthnet/ui/app.py

@@ -289,16 +289,21 @@ class UiApp:
                     build_marketplace_tab(self._bus)
                 with gr.Tab("Files"):
                     build_files_tab(self._bus)
-                with gr.Tab("🔬 Nemotron"):
-                    build_nemotron_tab(self._bus)
-                with gr.Tab("🎙 Voice"):
-                    build_voice_tab(self._bus)
-                with gr.Tab("🖼 Image"):
-                    build_image_tab(self._bus)
-                with gr.Tab("📄 OCR"):
-                    build_ocr_tab(self._bus)
-                with gr.Tab("🌍 Translation"):
-                    build_translation_tab(self._bus)
+                # [Disabled: event loop error in HF Spaces worker thread]
+                # with gr.Tab("🔬 Nemotron"):
+                #     build_nemotron_tab(self._bus)
+                # [Disabled: transcript backend unavailable]
+                # with gr.Tab("🎙 Voice"):
+                #     build_voice_tab(self._bus)
+                # [Disabled: Florence2 forced_bos_token_id config error]
+                # with gr.Tab("🖼 Image"):
+                #     build_image_tab(self._bus)
+                # [Disabled: TrOCR model compatibility issue]
+                # with gr.Tab("📄 OCR"):
+                #     build_ocr_tab(self._bus)
+                # [Disabled: translation backend not configured]
+                # with gr.Tab("🌍 Translation"):
+                #     build_translation_tab(self._bus)
                 with gr.Tab("Emergency"):
                     build_emergency_tab(self._bus, self._state_bus)
                 with gr.Tab("Settings"):

hearthnet/ui/tabs/ask.py

@@ -8,6 +8,12 @@ The routing trace shows exactly which node answered and why.
 No hardcoded responses. If no LLM is configured, an UnavailableBackend
 error is surfaced directly rather than fabricating an answer.
 
+LLM Models:
+- MiniCPM3-4B (OpenBMB default) requires trust_remote_code=True when loading via
+  transformers.from_pretrained() — the model repo contains custom modeling code.
+  HF Transformers backend (app.py) passes this flag; local-first vLLM/llama.cpp
+  endpoints do not need it (they handle the model internally).
+
 Spec: docs/M04-llm.md, docs/M05-rag.md, docs/M03-bus.md §4
 """