Voice mode should use the selected Assistant's persona/system prompt

[philmerrell]

Problem

Voice mode ignores the active Assistant entirely. A faculty member using a role-play Assistant (AI embodies a character) found voice mode drops the persona and behaves like a generic feedback chatbot — no amount of Assistant prompt tuning can fix it, because the persona is never sent.

Current behavior

• voice_routes.py (backend/src/apis/inference_api/chat/voice_routes.py:370-375) constructs VoiceAgent with only session_id, user_id, auth_token, enabled_tools — no system_prompt and no assistant id.
• With no system_prompt, BaseAgent.__init__ (backend/src/agents/main_agent/base_agent.py:132-137) builds the generic default prompt.
• VoiceAgent._build_voice_system_prompt() (voice_agent.py:140-150) appends a fixed voice addendum (concise, conversational, etc.) to that default.
• The frontend WebSocket config frame only carries session_id (voice-chat.service.ts:302-306), so there's nothing server-side to resolve an assistant from.

Voice does load prior text history for continuity (_load_text_history), but that's conversation messages, not the persona.

Proposed change

• Frontend: include the active assistant/agent id in the voice config frame.
• Backend: resolve the assistant's system prompt (with the same access checks the text path uses) and pass it into VoiceAgent, then append the voice addendum on top of it.
• Consider how the voice addendum interacts with strongly-in-character personas (e.g. "confirm understanding before taking actions" may fight a role-play prompt — the addendum may need to be softened or made persona-aware).

Notes

Driving use case: instructor role-play scenarios used in courses. Related: voice selection is tracked separately.