Use Logfire managed variables for prompt instructions

[dmontagu]

Summary

• move the agent instruction string into a Logfire managed variable with the bad demo prompt as the code default
• resolve instructions at runtime via @relations_agent.instructions so the same code path works for local GEPA runs and a deployed server
• switch eval, golden generation, and GEPA candidate testing to update the local Logfire variable provider instead of overriding agent instructions directly

Validation

• uv run ruff check task.py main.py cases.py adapter.py README.md
• uv run basedpyright task.py main.py cases.py adapter.py
• managed-variable GEPA check with openai:gpt-4.1-mini task model and gateway/anthropic:claude-opus-4-6 proposer on the existing golden dataset:
  • baseline held-out test score on 25 cases: 69.04%
  • GEPA validation score on 8-case val slice: 57.31% -> 82.31%
  • optimized held-out test score on the same 25 cases: 81.90%

Notes

• the CLI now uses LocalVariablesOptions so GEPA can mutate the same variable-backed instruction path locally
• a deployed server can switch to remote Logfire managed variables without changing the agent instruction code path