feat(headless): add resumable GLM harness benchmark

[Astro-Han]

Summary

• Add a fixed, resumable Terminal-Bench 2.1 comparison between Maka and OpenCode on GLM-5.2 Max.
• Freeze task order, harness identities, reasoning effort, pricing, and execution constraints in a fingerprinted manifest.
• Export separate Pass@1 effectiveness and cache-aware API-equivalent economy reports in JSON, CSV, and Markdown.

Why

We need reproducible evidence for comparing harness effectiveness and economy without mixing model, prompt, task-order, or billing changes into the result. This extends the existing headless, fixed-prompt, and Harbor seams so a 40-task pilot can resume into the full 89-task suite without rerunning completed cells.

Closes #861.

Scope

This PR adds benchmark infrastructure only. It does not run the benchmark, commit result artifacts, add Kimi or Pi arms, introduce LiteLLM, or change product defaults.

Verification

• npm test --workspace @maka/headless — 822 tests passed after rebasing onto the latest origin/main.
• npm run check:stale — passed.
• git diff --check origin/main...HEAD — passed.
• Secret-shaped token scan over the branch diff — no matches.
• No model or benchmark calls were made.
• No visual evidence is applicable because this change has no user-visible UI.

Impact

Headless benchmark operators gain a single fixed entrypoint with append-only resume state, execution attestation, and normalized reports. Existing product behavior and public defaults are unchanged; no migration is required.

Reviewer notes

Please focus on fairness invariants, resume correctness, secret isolation, OpenCode token accounting, and whether the implementation reuses the closest existing headless/Harbor seams without parallel state.

Ready for review

• Verification is complete and the results above are current
• Impact, compatibility, migration, documentation, and release-note needs are addressed
• User-visible UI/UX changes include visual evidence, or Verification explains why it is not applicable