feat(sandbox): add isolated two-phase evaluation and Meta-ReMe workspace tooling

[xyf2020]

Summary

This PR introduces the sandbox and workspace foundations required to evaluate and optimize ReMe memory implementations reproducibly.

It adds an isolated Docker-based benchmark environment, a two-phase build/query evaluation workflow, and the initial Meta-ReMe data preparation layer. Together, these changes make it possible to construct memory once, restart the application at a persistence boundary, evaluate multiple queries against the resulting workspace, and export structured artifacts for offline analysis.

What changed

• Added DockerReMeSandboxFactory and DockerReMeSandbox for running ReMe candidates inside isolated AgentScope Docker workspaces.
• Added support for both uploaded source snapshots and candidates preinstalled in Docker images.
• Introduced batch-oriented run_build() and run_queries() APIs:
  • build jobs share one Application instance;
  • query evaluation starts a fresh Application and reuses it sequentially;
  • individual query failures are recorded without stopping later queries;
  • scores, token deltas, raw results, logs, and errors are written to stable JSON artifacts.
• Added selective export profiles for analysis, evaluation, full case state, and reusable workspace snapshots.
• Added workspace upload, container reuse, serialized operations, safe archive handling, and explicit Git checkpoints for memory history.
• Added configurable exact-path logging so query-specific logs can be captured without unrelated startup or shutdown output.
• Added Docker build definitions and packaging extras for host-side sandbox control and container runtime dependencies.
• Added English and Chinese sandbox documentation covering isolation, lifecycle, artifact contracts, and usage examples.

Meta-ReMe foundation

The new meta-reme modules provide strict Pydantic contracts for datasets, validation selections, fingerprints, budgets, harnesses, and persisted run state. YAML-driven entry points now normalize LongMemEval and BEAM training data into a common case/session/query representation, create a locked file-native workspace, and install the normalized search dataset as read-only data.

A deterministic bundle builder creates minimal default, lme, or beam ReMe distributions from an explicit manifest. The generated bundle is validated and initialized as a local Git repository with a stable baseline commit, preparing it for later harness proposal and evaluation stages.

Compatibility and safety

Existing single-job sandbox helpers remain available. Candidate snapshots exclude common credentials, caches, datasets, runtime state, and VCS metadata, while archive upload validates paths, duplicates, symbolic links, and special files before extraction.

Docker integration and credential-dependent benchmark tests were not run locally.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ xyf2020
✅ jinliyl
❌ sa-buc

---

sa-buc seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}