Phase 3B: Implement pluggable LLM client for summary generation

- Create llm_client.py with 5 provider implementations (Anthropic, OpenAI, Ollama, OpenRouter, Google)
- Add build_prompt() helper to construct system/user prompts from templates
- Wire up POST /transcribe/sessions/{id}/summaries endpoint to call LLM client
- Return generated content + token counts (input_tokens, output_tokens)
- API keys passed per-request, never stored or logged
- Uses prompt templates from prompts.py based on summary_type

routers/transcribe/sessions.py

@@ -16,6 +16,10 @@ from modules.transcription.models import (
     SummaryResponse,
     ExportFormat,
 )
+from modules.transcription.llm_client import call_llm, build_prompt
+import logging
+
+logger = logging.getLogger(__name__)
 
 router = APIRouter()
 
@@ -211,28 +215,86 @@ async def generate_summary(
     summary_request: SummaryGenerateRequest,
     user_id: str = Depends(get_user_id),
 ):
-    """Generate a summary for a session (Phase 1 stub)."""
+    """Generate a summary for a session using the specified LLM provider.
+
+    Phase 3: Full implementation — calls the pluggable LLM client with
+    prompt templates from prompts.py. API key is passed per-request and
+    never stored or logged.
+    """
     supabase = get_supabase_client()
     
     # Verify session exists and user owns it
-    session_check = supabase.supabase.table("transcription_sessions").select("id").eq("id", session_id).eq("user_id", user_id).execute()
+    session_check = supabase.supabase.table("transcription_sessions").select("*").eq("id", session_id).eq("user_id", user_id).execute()
     
     if not session_check.data:
         raise HTTPException(status_code=404, detail="Session not found")
     
-    # Phase 1 stub: TODO implement LLM call in Phase 3
-    content = "[TODO: Generate summary via LLM — provider={}, model={}]".format(
-        summary_request.provider, summary_request.model
-    )
+    session = session_check.data[0]
+    
+    # Build transcript from segments (or use segment range)
+    segments_query = supabase.supabase.table("transcription_segments").select("*").eq("session_id", session_id).order("sequence_index")
+    segments_result = segments_query.execute()
+    
+    if not segments_result.data:
+        raise HTTPException(status_code=400, detail="No segments found for this session")
+    
+    # Apply segment range filter if specified
+    segments = segments_result.data
+    if summary_request.segment_range and len(summary_request.segment_range) == 2:
+        start_idx, end_idx = summary_request.segment_range[0], summary_request.segment_range[1]
+        if start_idx is not None and end_idx is not None:
+            segments = segments[start_idx:end_idx]
+        elif start_idx is not None:
+            segments = segments[start_idx:]
+        elif end_idx is not None:
+            segments = segments[:end_idx]
+    
+    # Build full transcript text from segments
+    transcript_parts = [s["text"] for s in segments if s.get("text")]
+    transcript = "\n".join(transcript_parts)
+    
+    if not transcript.strip():
+        raise HTTPException(status_code=400, detail="Transcript is empty — cannot generate summary")
+    
+    # Build prompt from template
+    system_prompt, user_message = build_prompt(summary_request.summary_type, transcript)
+    
+    # Call the LLM client
+    try:
+        llm_result = await call_llm(
+            provider=summary_request.provider,
+            model=summary_request.model,
+            api_key=summary_request.api_key,
+            system_prompt=system_prompt,
+            user_message=user_message,
+        )
+    except Exception as e:
+        logger.error(f"LLM call failed: {e}")
+        raise HTTPException(status_code=502, detail=f"LLM generation failed: {str(e)}")
+    
+    # Determine segment range for storage
+    seg_start = None
+    seg_end = None
+    if summary_request.segment_range and len(summary_request.segment_range) == 2:
+        seg_start = summary_request.segment_range[0]
+        seg_end = summary_request.segment_range[1]
+    
+    # Build the prompt that was used (for audit trail)
+    prompt_used = f"{system_prompt}\n\n{user_message}" if summary_request.summary_type != "segment" else user_message
     
     # Save summary to database
     summary_data = {
         "session_id": session_id,
         "user_id": user_id,
         "summary_type": summary_request.summary_type,
-        "content": content,
+        "content": llm_result.content,
+        "prompt_used": prompt_used,
         "llm_provider": summary_request.provider,
         "llm_model": summary_request.model,
+        "input_tokens": llm_result.input_tokens,
+        "output_tokens": llm_result.output_tokens,
+        "segment_range_start": seg_start,
+        "segment_range_end": seg_end,
     }
     
     result = supabase.supabase.table("transcription_summaries").insert(summary_data).execute()