refactor(llm): converge service-level LLM calls to unified client

[Congregalis]

Summary

This PR converges remaining service-level direct LLM HTTP calls into the unified LLM client abstraction.

It is a preparatory refactor to reduce protocol drift and make future provider expansion (e.g., OpenAI Responses API, Gemini) safer and more localized.

Closes #60

What Changed

• Replaced direct /chat/completions HTTP call path in agent-to-agent messaging with create_llm_client() + unified LLMMessage/LLMResponse flow.
• Replaced direct /chat/completions HTTP call path in supervision reminder reply generation with create_llm_client() + unified flow.
• Kept existing behavior intent (tool loop, token accounting, message flow, limits) as close as possible.
• Ensured client lifecycle is explicitly closed in refactored paths.

Why

Previously, these service paths bypassed the unified client and hardcoded request/response assumptions. That made future provider support harder and increased inconsistency risk across runtime flows.

After this change, protocol-specific handling is more centralized in the LLM client layer, which lowers the cost of adding provider types like OpenAI Responses and Gemini.

Scope Notes

• This PR does not introduce new providers.
• This PR does not perform a model schema redesign (provider + protocol).
• This PR focuses on convergence with minimal behavior disruption.

Impact

• Feature-level behavior should remain consistent for current providers.
• Future provider onboarding should require fewer service-level edits.

Validation

• Verified modified Python modules compile successfully.
• Confirmed service-level hardcoded /chat/completions usage was removed from the affected paths.