refactor(stage-3): add Keep-Alive connection pool for Ollama, implement offline lemma resolution in dictionary

app/main.py

@@ -36,7 +36,7 @@ from app.reader import (
     fetch_and_parse_rfc, parse_rfc_text, fetch_and_parse_wikipedia,
     extract_wikipedia_title_and_lang
 )
-from app.ai import translate_sentence_ai, get_ollama_status, warmup_ollama_model
+from app.ai import translate_sentence_ai, get_ollama_status, warmup_ollama_model, close_ollama_client
 from app.phrasal import detect_phrasal_verb
 from app.cefr import (
     get_cefr_level_summaries,
@@ -97,6 +97,7 @@ async def lifespan(app: FastAPI):
     asyncio.create_task(prewarm_all_listening_audio())
     yield
     await close_http_client()
+    await close_ollama_client()
 
 
 app = FastAPI(