fix: apply SSE timeout to HTTP client and fix mutable default argument in llms.py

[amathxbt]

Bugs

Two bugs in demo/nesa/backend/llms.py:

Bug 1 — SSE timeout silently ignored

sse_message_handler accepts a timeout parameter (default 60s) but passes timeout=None to the httpx client:

# Before (broken) — connection can hang forever
async with client.stream(
    "POST", ..., timeout=None  # ignores the timeout parameter!
) as response:

Impact: SSE connections to the inference backend can hang indefinitely regardless of the timeout argument. This blocks the event loop and can deadlock the server under network failures.

Bug 2 — Mutable default argument

perform_inference uses a mutable list as a default value:

# Before (broken)
history: Optional[List[str]] = [],  # shared across all calls!

Impact: Python creates this list object once at function-definition time. If any caller mutates history in place, the mutation persists across subsequent calls — a classic Python footgun that can cause cross-session state corruption.

Fix

1. Pass the actual timeout argument to the httpx stream call.
2. Changed default to None with an or [] guard at the call site.

# After (fixed)
async with client.stream("POST", ..., timeout=timeout) as response:

# and
history: Optional[List[str]] = None,
...
prompt_template = generate_prompt_template(..., history=history or [])