feat: define Autoresearch metric-experiment preset

[devkade]

Summary

Define Autoresearch as a metric-experiment preset in the RunContract Harness.

Parent roadmap: #114
Track: C - Preset evolution
Related: #118, #167, #172

Problem

Autoresearch is the bounded experiment/optimization preset, but its RunContract-facing contract is not explicit. The harness needs a clear preset boundary for metric selection, benchmark execution, experiment ledger evidence, stopping rules, and anti-Goodhart safeguards.

Scope

• Define the Autoresearch metric-experiment preset contract.
• Specify required inputs: goal, benchmark or metric target, constraints, evidence standard, stop conditions, and acceptable trade-offs.
• Specify required artifacts such as contract, benchmark/checks, experiment ledger, ideas, decision report, and verification report.
• Define how objective/evaluation signals can inform Autoresearch without becoming hidden hard-blocking authority.
• Identify follow-up implementation slices if existing Autoresearch behavior needs alignment.

Non-goals

• No neural/RL learning implementation.
• No unbounded autonomous research loop.
• No score-only completion authority.
• No runtime plugin/module retirement behavior.
• No command rename or storage-root migration.

Acceptance criteria

• Autoresearch preset contract is documented or implemented with tests.
• Metric, benchmark, guardrail, and stop-condition expectations are explicit.
• Experiment ledger evidence requirements are defined.
• Anti-Goodhart safeguards reference the harness evaluator policy.
• Objective/evaluation signals remain advisory unless a separate issue authorizes stronger gates.

Verification

• Design review against Kapi → Ilchul RunContract Harness roadmap #114 RunContract boundary.
• Review against docs/runcontract-harness-evaluator.md.
• Relevant Autoresearch tests if implementation changes are included.