Tune general task execution guidance

[ericleepi314]

Summary

• prioritize evidence-ranked, low-cost diagnostic checks before broad exploration
• use proportionate verification and stop after requirements are demonstrably satisfied
• avoid redundant temporary tests and repeated passing checks
• group related mechanical edits while keeping unrelated changes separate

The guidance is model-, provider-, and benchmark-agnostic.

Validation

• pytest -q tests/test_system_prompt_full.py tests/test_live_base_system_prompt.py tests/test_prompt_assembly.py (82 passed)
• git diff --check