feat: add main harness core

[KumaKuma2002]

Summary

• Add a self-contained main_harness/ module for PawBench harness-core research and evaluation.
• Include the AgentScope candidate implementation with feature switches, runtime isolation, tracing, verification, and attribution support.
• Add a governed Feature-development handoff: an accepted H-to-F request is implemented by Qwen3.8-Max through Claude Code with a repository-local skill, strict MCP/tool policy, isolated credential environment, and independent admission validation.
• Normalize active Feature display names so slash-separated labels are not part of the current contract.

Why

This provides an auditable harness-core implementation for studying which harness features affect agent outcomes. Feature selection, implementation, receipt validation, activation, and paired evaluation remain separate, so the coding agent cannot silently promote a Feature.

Scope and impact

• All changes remain under main_harness/.
• The Qwen route reads only the DashScope configuration at runtime; no credential, run record, cache, archive, dataset, or machine-specific path is included.
• Existing PawBench runtime paths remain unchanged until a caller explicitly activates an admitted Feature.
• The installed development command can render --help outside a checkout, while prepare/run/verify deliberately require an explicit PawBench --workspace-root for the canonical manifest and local skill.

Validation

• 113 passed, 6 skipped — full main_harness pytest suite.
• 43 passed — AgentScope candidate suite, including 10 Feature-development contract tests.
• Built the AgentScope wheel and ran its console entrypoint from a directory outside the checkout.
• Confirmed the configured DashScope model listing exposes qwen3.8-max without running a mutable coding-agent pass.

[helloml0326]

1.是否把AgentScope封装为Harbor的Agent，可以通过HarborBridgeAgent方式调用
2.当前AgentScope可编程部分是否满足pawbench 2.0需要的Feature
3.整体代码定义比较多，是否可以把归因->Feature->效果验证这个pipeline建模，比如一次效果验证是一个 experiment，包含多个验证对象，每个验证对象包含(归因、feature AgentScope对应的模块等)