Add AutoR agent preset

[tangxiangru]

Adds a built-in preset for AutoR, following the Add Your Own Agent instructions.

What AutoR is

A stage-structured research harness that drives Claude Code or Codex CLI through an eight-stage pipeline (literature survey → hypothesis → study design → implementation → experimentation → analysis → writing → dissemination). Each stage boundary is normally a human approval gate; for benchmark runs a reviewer agent takes that role, so a run needs no human at any point.

Changes

• evaluation/agents.json — the autor preset.
• evaluation/static/logos/autor.svg — 64×64, matching the existing rh.svg house style.
• README.md — the installation table, the supported-agents table, and the prose list of built-in agents.

"autor": {
  "label": "AutoR",
  "icon": "A",
  "logo": "/static/logos/autor.svg",
  "cmd": "python3 /abs/path/to/AutoR/rcb_agent.py --workspace <WORKSPACE> --prompt <PROMPT>"
}

AutoR is a repository checkout rather than an installed CLI, so the preset uses the /abs/path/to/ placeholder convention already used by the ResearchHarness entry in the same file.

Testing

Driven through TaskRunner with the preset exactly as written (with only the /abs/path/to/ prefix substituted for a real checkout):

preset loaded: AutoR | /static/logos/autor.svg
workspace: workspaces/Chemistry_000_20260806_035047
STATUS: completed | exit_code: 0
report.md: 3185 bytes

The run produced report/report.md at the benchmark path along with report/images/, code/ and outputs/, left data/ and related_work/ untouched, and streamed JSON lines to stdout so _agent_output.jsonl and the run browser's model detection both work. _meta.json came back completed from the harness as expected.

Notes

• AutoR never reads stdin in this mode. An interactive prompt reachable during a benchmark run raises rather than blocking, which matters here because the harness passes the launching terminal's stdin through to the agent.
• Setup for the AutoR side is documented at docs/researchclawbench.md.
• This PR adds the preset only; no leaderboard results are included.

Happy to adjust the label, icon, logo, or command shape to whatever you prefer.