[Fix] Quick start does not work verbatim, and operational noise #120

• README/README_ko/guide say python -m api run server after uv sync; without an activated venv that is the system Python (ModuleNotFoundError). Use uv run --extra torch python -m api run server or add a script entry point.
• print_output=True (server.py) prints every prompt and answer; models/qwen3/model.py prints the system prompt at import; spinner output in models/base.py. Use logging.
• Sessions never expire (settings.py).
• web_search.py duplicates load_dotenv() and its fallback block and imports requests, which is not a declared dependency.
• Engines are not closed on shutdown (no lifespan hook); Ctrl-C with open streams is untested.

Done when: a CI job runs the README quick-start commands literally against a fake/small model; logging replaces prints; sessions expire after a configurable idle time; requests is declared or dropped; a shutdown test exits within N seconds with an open stream.

Found in the 2026-10-03 usability audit (code reading, not run).