perf: preload all models at startup, improve Q&A context and auto-submit

- Preload Whisper-small, Qwen2.5-3B-Instruct, and TTS at startup
- Use .change() event for auto-submit (fires after audio value committed)
- Pass full story context (or top-5 relevant paragraphs) to Q&A
- Fix total_mem -> total_memory attribute error

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

app.py

@@ -387,11 +387,20 @@ else:
     _default_voice_name = None
     print("[MomsVoice] No saved voice profile — using stock voice as default.")
 
-# Preload the 0.6B CustomVoice model at startup for fast first-play response
+# Preload models at startup for fast response
 print("[MomsVoice] Preloading Qwen3-TTS 0.6B CustomVoice model...")
 from voice_clone import get_custom_voice_model
 get_custom_voice_model()
-print("[MomsVoice] Model ready.")
+print("[MomsVoice] TTS model ready.")
+
+print("[MomsVoice] Preloading Whisper-small ASR model...")
+from inference import get_asr_pipeline, get_qa_model
+get_asr_pipeline()
+print("[MomsVoice] ASR model ready.")
+
+print("[MomsVoice] Preloading Qwen2.5-3B-Instruct Q&A model...")
+get_qa_model()
+print("[MomsVoice] Q&A model ready. All models preloaded.")
 
 
 # Gradio Application Core setup
@@ -1022,10 +1031,18 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
         outputs=[answer_display, answer_audio]
     )
 
-    # Auto-submit when voice recording stops (end-of-question detection)
-    question_audio.stop_recording(
-        handle_question_submit,
-        inputs=[question_text, question_audio, paragraphs_state, timeline_slider, voice_profile_state],
+    # Auto-submit when voice recording completes
+    # Use .change() instead of .stop_recording() — .change() fires after audio value is committed
+    def auto_submit_on_recording(audio_path, question_txt, paragraphs, slider_val, profile_id):
+        """Auto-submit only when a new recording is available (not on clear)."""
+        if not audio_path:
+            # Audio was cleared, don't submit
+            return gr.update(), gr.update()
+        return handle_question_submit(question_txt, audio_path, paragraphs, slider_val, profile_id)
+
+    question_audio.change(
+        auto_submit_on_recording,
+        inputs=[question_audio, question_text, paragraphs_state, timeline_slider, voice_profile_state],
         outputs=[answer_display, answer_audio]
     )
 

inference.py

@@ -70,7 +70,7 @@ def get_qa_model():
         # Check available VRAM — if less than 3GB free, use CPU
         use_gpu = False
         if torch.cuda.is_available():
-            free_vram = torch.cuda.get_device_properties(0).total_mem - torch.cuda.memory_allocated(0)
+            free_vram = torch.cuda.get_device_properties(0).total_memory - torch.cuda.memory_allocated(0)
             free_vram_gb = free_vram / (1024**3)
             logger.info("Free VRAM: %.1f GB", free_vram_gb)
             if free_vram_gb >= 3.0:
@@ -98,11 +98,19 @@ def get_qa_model():
 
 
 def _get_relevant_context(paragraphs: list[str], current_idx: int, question: str) -> str:
-    """Simple TF-IDF-like retrieval: return the most relevant paragraphs as context."""
+    """Return full story with emphasis on current section for context."""
     if not paragraphs:
         return ""
 
-    # Use paragraphs around the current position + simple keyword overlap
+    # Build context: full story (truncated if too long) with current paragraph highlighted
+    total_text = "\n\n".join(paragraphs)
+
+    # If story is short enough (< 2000 chars), use it all
+    if len(total_text) <= 2000:
+        current_marker = f"\n\n[Currently reading]: {paragraphs[current_idx]}" if current_idx < len(paragraphs) else ""
+        return total_text + current_marker
+
+    # For longer stories: use top relevant paragraphs + surrounding context
     question_words = set(question.lower().split())
 
     scored = []
@@ -110,13 +118,19 @@ def _get_relevant_context(paragraphs: list[str], current_idx: int, question: str
         para_words = set(para.lower().split())
         overlap = len(question_words & para_words)
         # Boost paragraphs near current position
-        proximity_bonus = max(0, 3 - abs(i - current_idx))
+        proximity_bonus = max(0, 5 - abs(i - current_idx))
         scored.append((overlap + proximity_bonus, i, para))
 
     scored.sort(key=lambda x: x[0], reverse=True)
-    # Take top 3 most relevant paragraphs
-    top_paras = [s[2] for s in scored[:3]]
-    return "\n\n".join(top_paras)
+    # Take top 5 most relevant paragraphs
+    top_paras = sorted(scored[:5], key=lambda x: x[1])  # sort by position
+    context = "\n\n".join(s[2] for s in top_paras)
+
+    # Add current paragraph marker
+    if current_idx < len(paragraphs):
+        context += f"\n\n[Currently reading]: {paragraphs[current_idx]}"
+
+    return context
 
 
 def answer_story_question(