fix(suggestion_handler): split system+user so non-OpenAI LLMs accept the request

[Ptah-CT]

Fixes #1527.

Problem

handle_get_suggestion_queries builds a single system message:

message_list = [{"role": "system", "content": suggestion_prompt.format(memories=memories)}]

OpenAI accepts this. MiniMax (both Text API and Anthropic-compatible API) and Anthropic itself reject it with HTTP 400 (chat content is empty (2013) / messages must not be empty (2013)). Per MiniMax's own minimal example in the docs, every request must contain at least one user turn.

Change

Split into a short system persona + user message:

message_list = [
    {"role": "system",
     "content": "You generate suggestion queries based on the user's recent memories."},
    {"role": "user",
     "content": suggestion_prompt.format(memories=memories)},
]

This is a strict widening:

• OpenAI: still accepted
• MiniMax (Text + Anthropic-compat): now accepted
• Anthropic native: now accepted

No behaviour change for existing OpenAI deployments.

Type

• Bug fix (non-breaking change which fixes an issue)

Tested

• POST /product/suggestions with MOS_CHAT_MODEL=MiniMax-M2.7, OPENAI_API_BASE=https://api.minimax.io/v1:
  • Before: 500 (or 'NoneType has no attribute replace' if running without fix(utils): re-raise in timed_with_status when no fallback configured #1524).
  • After: 200 with {\"data\":{\"query\":[\"...\",\"...\",\"...\"]}} populated by MiniMax-M2.7.
• Verified the same fix path works end-to-end (LLM call → JSON parse → SuggestionResponse).

Checklist

• Self-review
• Comment added at the message-list construction explaining why both roles are needed
• Unit test (would need an LLM mock; happy to add if maintainers want one)
• Linked to issue
• Docs update — should we mention this in any provider-compatibility table? Could open a follow-up doc PR.