Daria 2.3.0

v2.3.0: Qwen 3.8 coding, prompt-cache reuse, and speculative decode

Bring the coding agent in line with upstream Chaty 2.1.7 without TikTok
or extra UI languages. Tool results use the tool role, reasoning stays
in its own field, and per-turn context no longer busts the KV cache —
so Qwen 3.8 long runs resume instead of re-reading the conversation.
llama.cpp moves to a Muse-Glimmer build with optional MTP speculative
decoding. Browser clicks land on the named control, generations survive
a webview reload, and the system prompt no longer embeds the clock.