Reduce terminal noise during benchmark runs

[MaximilianAnzinger]

Summary

• Remove per-metric insights from terminal output — the detailed qualitative reasoning (which can produce thousands of lines) no longer floods the CLI. It is still written to summary.txt in each run's result directory.
• Move compliance diagnostics behind --debug — difficulty band checks and instruction adjustment logs in base.py and benchmark.py are now logger.debug() instead of print(), so they appear only when running with --debug (and always in run.log).
• Keep what matters on screen — per-metric progress lines and the final stats summary table remain unchanged.

Test plan

• All 199 tests pass
• Run a benchmark — terminal shows only progress + stats table
• Verify summary.txt still contains full insights
• Verify --debug surfaces compliance diagnostics

🤖 Generated with Claude Code