Add safe prompt framing for retrieved conversation memory

[lbedner]

Summary

Ensure retrieved memory is framed, timestamped, and has instructions to prefer recent conversation if conflicts exist.

Acceptance Criteria

• Retrieved hits inserted in concise, framed block with role and timestamp
• System prompt contains explicit instructions about precedence and uncertainty handling
• Config values for characters-per-hit cap (trim long hits)
• Document the prompt format and an example in repo docs

Implementation Notes

• Add constants for prompt snippet formatting in AI service or prompts module
• Example framing:

  Retrieved memory (may be incomplete):
  - [2026-01-02 12:34] USER: ...
  - [2026-01-02 12:35] ASSISTANT: ...
  

• System prompt rules:
  • "If retrieved memory conflicts with recent conversation, prefer recent."
  • "If memory is unclear, ask a clarifying question."

Files to Change

• app/services/ai/prompts.py - New/edited prompt templates
• app/cli/docs or README.md - Documentation
• Tests for prompt formatting

Dependencies

• Requires Retrieve relevant conversation history snippets before prompt construction #420 (Retrieval step) to be completed first

Estimate

4-8 hours