examples: Streamlit chat over the Gemini proxy demo

[Ar9av]

examples/gemini-proxy-demo/demo.py (#388) shows the proxy governing Google Gen AI against a stub upstream, offline and keyless. This adds the other half: a real chat UI on the real API, where the only Prismor-aware line is the base_url on the client.

client = genai.Client(
    api_key=os.environ["GEMINI_API_KEY"],
    http_options=types.HttpOptions(base_url=os.getenv("PRISMOR_PROXY", "http://127.0.0.1:7080")),
)

22 lines total, including the chat loop.

Verified against prismor 1.50.0

check	result
streams through the proxy in enforce mode	PONG
prompt screening reaches the wire	credential pasted into chat reached Gemini as @@SECRET:...@@
overhead vs direct call	2.0-2.5s direct, 2.6-4.5s proxied (n=4)

Two things worth knowing

Port 7080 collides. It is both the proxy default and demo.py's hardcoded port. A second listener binds alongside an existing one rather than failing, and requests then land on whichever, under the wrong workspace and policy. This cost real debugging time on a box already running a long-lived proxy under a LaunchAgent. The README says to check lsof and pass --port; making demo.py's port overridable is a reasonable follow-up, left out here to keep the diff to new files.

gemini-2.5-flash is retired — the API returns 404 pointing new users at gemini-3.6-flash, which is what this uses.