chore: delete eval traces, run artifacts, and dead code

- eval_traces/: 741MB / 339 LFS files of SWE-Atlas benchmark traces. Not
  runnable examples; belongs in a release artifact, not a public examples repo.
- review_service/: 145 LOC that could not build (Dockerfile copied
  evolver/seed_organism_v2.json, which does not exist) and duplicated
  github_app's flow behind a different trigger.
- pr_review_agent/output/ (64 run artifacts) and plan.md: both already listed
  in .gitignore, tracked only because they predate the rule.
- pr_review_agent/{research_notes,research_findings,strategy_research,eval_results}.md:
  1094 lines of agent working notes with stale Feb-2026 leaderboards.
- pr_review_agent/{evolve_prompt.py,evolver/{run,mutator,evaluator,failure_case}.py}:
  ~1250 LOC importing darwinian_evolver, a module that does not exist.
  evolver/organism.py stays — review.py depends on it.
- pr_review_agent/{evaluate,online_eval}.py: benchmark harness, never shipped.
- .claude/ralph-loop.local.md: live autonomous-loop state (active: true).
- .gitattributes: the 6 LFS patterns only covered eval_traces/.