Enforced locked backend config for BrainForge personas

[NeonCharlie-24]

Description

BrainForge advisors advertise backend_locked: true but the flag was ignored during LLM routing, allowing locked advisors to run against user-selected backends (Gemini, vLLM).

This PR enforces backend_locked end-to-end:

• Backend: resolve_llm_clients() preserves the locked persona's own client in both uniform and hybrid modes. switch_provider rejects explicit overrides for locked personas. /api/config derives the flag from the model instead of hardcoding it.
• Frontend: Config draft excludes locked personas so they never appear in the save payload. Submit handler sends mode: "uniform" or "hybrid" based on which overrides exist ("uniform" mode was previously never sent by the frontend).

Issues

Closes #86

Other Notes

Manual E2E

1. Add this line in chat.py → resolve_llm_clients(), after each persona_clients[pid] = ... assignment (uniform branch, line ~49; hybrid branch, line ~62):

logger.info("resolve_llm_clients: %s -> %s (locked=%s)", pid, type(persona_clients[pid]).__name__, persona.backend_locked)

2. Rebuild the container and open backend logs
3. Open Settings --> Model and switch the model configuration
4. Start a new chat and review the logs to verfiy BrainForge personas set to BrainForge client

[NeonDaniel]

Changes look good; pulled locally to test and confirm edge case is resolved