codex-autoresearch adds autonomous research loop

This is less a coding assistant than a control system for disciplined, metric-driven experimentation. The value is in the guardrails: it turns “let Codex keep trying” into a reproducible loop that can actually converge.

• –One atomic change per iteration keeps the search space honest and makes rollback cheap.
• –Lessons persist across runs, so the system compounds knowledge instead of starting cold every time.
• –Parallel worktrees let it fan out hypotheses without mixing state, which matters for research-y codebases.
• –Best fit is measurable cleanup work: type reduction, test coverage, lint debt, performance tuning, and similar chores.
• –The big risk is obvious too: if the metric or guard is weak, the loop can optimize the wrong thing very efficiently.