Add MiniMax Code Harbor evaluation profile

[hetaoBackend]

Summary

• add a reproducible MiniMax Code profile that runs all 30 published task IDs through Harbor's mcode adapter
• replace MCode's networked per-task install with a checkpointed, SHA-256-verified Node/MCode bundle while preserving Harbor's normal configuration, execution, and trajectory paths
• pin Harbor, MiniMax Code, Node, Terminal-Bench, and DeepSWE inputs; overlay the repository's public task definitions onto the complete upstream task assets
• add Fireworks, Moonshot, OpenRouter, and Together Kimi K3 mappings, checkpoint/run identity checks, evidence collection, reporting metadata, and focused regression tests

Reproducibility pins

• Harbor 0.22.0
• @minimax-ai/code 0.2.7
• Node.js 22.23.2 (official Linux x64 archive SHA-256 verified)
• Terminal-Bench 2.1 dataset sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a
• DeepSWE commit 435ee89ec2f2e2289f33b0da4f992f0b7b7266b9

Every restored MCode checkpoint recomputes and compares the public-task tree, runnable Terminal-Bench and DeepSWE trees, MCode bundle, adapter source, and DeepSWE HEAD before a trial can run. An existing run ID also rejects changes to the checkpoint, harness/version, provider/model routes, timeout, or command identities.

Validation

• skills/frontierharness-eval/tests/test-minimax-code.sh
• bash -n skills/frontierharness-eval/scripts/*.sh skills/frontierharness-eval/tests/*.sh
• node --check skills/frontierharness-eval/scripts/*.mjs
• uvx ruff check skills/frontierharness-eval/agents/frontierharness_mcode.py
• skill metadata validation with Codex's quick_validate.py
• Harbor 0.22.0 resolution and full export of the pinned Terminal-Bench digest; verified all 21 selected Terminal-Bench and 9 DeepSWE overlays load from the resulting local task trees
• offline adapter install contract test and Harbor custom-agent --print-config validation

I did not run a real Runta/Kimi K3 task or the full 30-task sweep from this environment: it has no Runta/provider credentials and its Docker daemon is unavailable. This PR adds the reproducible integration workflow, not benchmark result claims.

Stack

This PR is intentionally based on and targets #1 (feat/third-part-harness), which introduces the agent-neutral evaluation workflow that this MiniMax Code profile extends.