fix: swap TASK_MODEL to Llama-3.1-8B-Instruct
Qwen/Qwen3-4B-Instruct-2507 is no longer routable on the HF inference
router, so resolveTaskModel() silently fell back to modelList[0]
(a reasoning model), and conversation titles came out as the model's
chain of thought ("The user wants a title for the chat th...").
Llama-3.1-8B-Instruct is available, follows the existing title prompt
cleanly across English/French/Chinese inputs, and emits no <think>
blocks — verified against the live HF router.
