feat(headless): add a resumable Maka vs OpenCode GLM-5.2 benchmark

[Astro-Han]

Goal

Produce reproducible evidence of whether Maka is more effective and token-efficient than OpenCode when both use GLM-5.2 Max under the same Terminal-Bench 2.1 conditions.

Scope

• Reuse the existing packages/headless Harbor path and Maka/OpenCode adapters.
• Run paired Pass@1 trials over a deterministic ordering of all 89 tasks.
• Treat the first 40 tasks as a resumable prefix of the full run.
• Freeze both harness configurations before the first formal trial.
• Store an immutable manifest and append-only results so interrupted runs execute only missing cells.
• Retry only infrastructure failures; agent failures remain benchmark outcomes.
• Report paired pass rates, token usage, and API-equivalent cost in JSON, CSV, and Markdown.
• Never store credentials or publish results automatically.

Done when

• Tests cover manifest identity, deterministic ordering, resume safety, failure classification, and metric calculation.
• A dataset-external smoke test validates both adapters without spending benchmark quota.
• The 40-task run can extend to all 89 tasks without rerunning valid results.
• The report supports an honest two-axis effectiveness/economy comparison.

Non-goals

• Pi, Kimi, Ollama, Pass@5, prompt tuning, LiteLLM, or broader model-catalog work.