Agent replies are no longer forced to follow the UI language

What changed

Agents now choose their reply language freely instead of always being told to respond in the UI language. A new Agent language setting (global default plus a per-agent override) lets users pin a specific reply language when they want one; when set, persona and workspace instructions can still override it.

Why this matters to the user

Previously every agent reply was forced into the UI language, which could fight with personas or instructions written in another language. After upgrading, agents follow the language of the conversation by default — usually what the user wants. Pinning a language will be possible once the new agent-language setting is exposed in the settings UI.

What the user should do

Nothing — automatic.

Notes for release manager

The global preference (agent.language) and the per-agent override are not yet exposed in the settings UI at the time of this PR; a follow-up adds the picker. Until then the default behavior (reply language follows the conversation) needs no configuration, so this entry stays notice. Changing an agent's language rebuilds its warm connection, so the first turn after a change pays full input-token cost until the new prompt prefix is cached.